llvm/test/CodeGen/AMDGPU/load-global-i32.ll

   1 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -mtriple=amdgcn -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=GCN-NOHSA -check-prefix=SI-NOHSA -check-prefix=FUNC %s
   2 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -mtriple=amdgcn--amdhsa -mcpu=kaveri -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=GCN-HSA -check-prefix=GCNX3-HSA -check-prefix=FUNC %s
   3 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -mtriple=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=GCN-NOHSA -check-prefix=GCNX3-NOHSA -check-prefix=FUNC %s
   4 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=r600 -mcpu=redwood < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=EG -check-prefix=FUNC %s
   5 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -mtriple=amdgcn--amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=GCN-HSA -check-prefix=FUNC %s
   6 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -mtriple=amdgcn--amdhsa -mcpu=gfx908 -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=GCN-HSA -check-prefix=FUNC %s
   7
   8 ; FUNC-LABEL: {{^}}global_load_i32:
   9 ; GCN-NOHSA: buffer_load_dword v{{[0-9]+}}
  10 ; GCN-HSA: {{flat|global}}_load_dword
  11
  12 ; EG: VTX_READ_32 T{{[0-9]+}}.X, T{{[0-9]+}}.X, 0
  13 define amdgpu_kernel void @global_load_i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
  14 entry:
  15   %ld = load i32, ptr addrspace(1) %in
  16   store i32 %ld, ptr addrspace(1) %out
  17   ret void
  18 }
  19
  20 ; FUNC-LABEL: {{^}}global_load_v2i32:
  21 ; GCN-NOHSA: buffer_load_dwordx2
  22 ; GCN-HSA: {{flat|global}}_load_dwordx2
  23
  24 ; EG: VTX_READ_64
  25 define amdgpu_kernel void @global_load_v2i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
  26 entry:
  27   %ld = load <2 x i32>, ptr addrspace(1) %in
  28   store <2 x i32> %ld, ptr addrspace(1) %out
  29   ret void
  30 }
  31
  32 ; FUNC-LABEL: {{^}}global_load_v3i32:
  33 ; SI-NOHSA: buffer_load_dwordx4
  34 ; GCNX3-NOHSA: buffer_load_dwordx3
  35 ; GCNX3-HSA: {{flat|global}}_load_dwordx3
  36
  37 ; EG: VTX_READ_128
  38 define amdgpu_kernel void @global_load_v3i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
  39 entry:
  40   %ld = load <3 x i32>, ptr addrspace(1) %in
  41   store <3 x i32> %ld, ptr addrspace(1) %out
  42   ret void
  43 }
  44
  45 ; FUNC-LABEL: {{^}}global_load_v4i32:
  46 ; GCN-NOHSA: buffer_load_dwordx4
  47 ; GCN-HSA: {{flat|global}}_load_dwordx4
  48
  49 ; EG: VTX_READ_128
  50 define amdgpu_kernel void @global_load_v4i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
  51 entry:
  52   %ld = load <4 x i32>, ptr addrspace(1) %in
  53   store <4 x i32> %ld, ptr addrspace(1) %out
  54   ret void
  55 }
  56
  57 ; FUNC-LABEL: {{^}}global_load_v8i32:
  58 ; GCN-NOHSA: buffer_load_dwordx4
  59 ; GCN-NOHSA: buffer_load_dwordx4
  60 ; GCN-HSA: {{flat|global}}_load_dwordx4
  61 ; GCN-HSA: {{flat|global}}_load_dwordx4
  62
  63 ; EG: VTX_READ_128
  64 ; EG: VTX_READ_128
  65 define amdgpu_kernel void @global_load_v8i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
  66 entry:
  67   %ld = load <8 x i32>, ptr addrspace(1) %in
  68   store <8 x i32> %ld, ptr addrspace(1) %out
  69   ret void
  70 }
  71
  72 ; FUNC-LABEL: {{^}}global_load_v9i32:
  73 ; GCN-NOHSA: buffer_load_dwordx4
  74 ; GCN-NOHSA: buffer_load_dwordx4
  75 ; GCN-NOHSA: buffer_load_dword
  76 ; GCN-HSA: {{flat|global}}_load_dwordx4
  77 ; GCN-HSA: {{flat|global}}_load_dwordx4
  78 ; GCN-HSA: {{flat|global}}_load_dword
  79 define amdgpu_kernel void @global_load_v9i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
  80 entry:
  81   %ld = load <9 x i32>, ptr addrspace(1) %in
  82   store <9 x i32> %ld, ptr addrspace(1) %out
  83   ret void
  84 }
  85
  86 ; FUNC-LABEL: {{^}}global_load_v10i32:
  87 ; GCN-NOHSA: buffer_load_dwordx4
  88 ; GCN-NOHSA: buffer_load_dwordx4
  89 ; GCN-NOHSA: buffer_load_dwordx2
  90 ; GCN-HSA: {{flat|global}}_load_dwordx4
  91 ; GCN-HSA: {{flat|global}}_load_dwordx4
  92 ; GCN-HSA: {{flat|global}}_load_dwordx2
  93 define amdgpu_kernel void @global_load_v10i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
  94 entry:
  95   %ld = load <10 x i32>, ptr addrspace(1) %in
  96   store <10 x i32> %ld, ptr addrspace(1) %out
  97   ret void
  98 }
  99
 100 ; FUNC-LABEL: {{^}}global_load_v11i32:
 101 ; SI-NOHSA: buffer_load_dwordx4
 102 ; SI-NOHSA: buffer_load_dwordx4
 103 ; SI-NOHSA: buffer_load_dwordx4
 104 ; GCNX3-NOHSA: buffer_load_dwordx4
 105 ; GCNX3-NOHSA: buffer_load_dwordx4
 106 ; GCNX3-NOHSA: buffer_load_dwordx3
 107 ; GCN-HSA: {{flat|global}}_load_dwordx4
 108 ; GCN-HSA: {{flat|global}}_load_dwordx4
 109 ; GCN-HSA: {{flat|global}}_load_dwordx3
 110 define amdgpu_kernel void @global_load_v11i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 111 entry:
 112   %ld = load <11 x i32>, ptr addrspace(1) %in
 113   store <11 x i32> %ld, ptr addrspace(1) %out
 114   ret void
 115 }
 116
 117
 118 ; FUNC-LABEL: {{^}}global_load_v12i32:
 119 ; GCN-NOHSA: buffer_load_dwordx4
 120 ; GCN-NOHSA: buffer_load_dwordx4
 121 ; GCN-NOHSA: buffer_load_dwordx4
 122 ; GCN-HSA: {{flat|global}}_load_dwordx4
 123 ; GCN-HSA: {{flat|global}}_load_dwordx4
 124 ; GCN-HSA: {{flat|global}}_load_dwordx4
 125 define amdgpu_kernel void @global_load_v12i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 126 entry:
 127   %ld = load <12 x i32>, ptr addrspace(1) %in
 128   store <12 x i32> %ld, ptr addrspace(1) %out
 129   ret void
 130 }
 131
 132 ; FUNC-LABEL: {{^}}global_load_v16i32:
 133 ; GCN-NOHSA: buffer_load_dwordx4
 134 ; GCN-NOHSA: buffer_load_dwordx4
 135 ; GCN-NOHSA: buffer_load_dwordx4
 136 ; GCN-NOHSA: buffer_load_dwordx4
 137
 138 ; GCN-HSA: {{flat|global}}_load_dwordx4
 139 ; GCN-HSA: {{flat|global}}_load_dwordx4
 140 ; GCN-HSA: {{flat|global}}_load_dwordx4
 141 ; GCN-HSA: {{flat|global}}_load_dwordx4
 142
 143 ; EG: VTX_READ_128
 144 ; EG: VTX_READ_128
 145 ; EG: VTX_READ_128
 146 ; EG: VTX_READ_128
 147 define amdgpu_kernel void @global_load_v16i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 148 entry:
 149   %ld = load <16 x i32>, ptr addrspace(1) %in
 150   store <16 x i32> %ld, ptr addrspace(1) %out
 151   ret void
 152 }
 153
 154 ; FUNC-LABEL: {{^}}global_zextload_i32_to_i64:
 155 ; GCN-NOHSA-DAG: buffer_load_dword v[[LO:[0-9]+]],
 156 ; GCN-HSA-DAG: {{flat|global}}_load_dword v[[LO:[0-9]+]],
 157 ; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0{{$}}
 158
 159 ; GCN-NOHSA: buffer_store_dwordx2 v[[[LO]]:[[HI]]]
 160 ; GCN-HSA: {{flat|global}}_store_dwordx2 v{{.+}}, v[[[LO]]:[[HI]]]
 161
 162 ; EG: MEM_RAT_CACHELESS STORE_RAW T{{[0-9]+}}.XY
 163 define amdgpu_kernel void @global_zextload_i32_to_i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 164   %ld = load i32, ptr addrspace(1) %in
 165   %ext = zext i32 %ld to i64
 166   store i64 %ext, ptr addrspace(1) %out
 167   ret void
 168 }
 169
 170 ; FUNC-LABEL: {{^}}global_sextload_i32_to_i64:
 171 ; GCN-NOHSA: buffer_load_dword v[[LO:[0-9]+]]
 172 ; GCN-HSA: {{flat|global}}_load_dword v[[LO:[0-9]+]]
 173 ; GCN: v_ashrrev_i32_e32 v[[HI:[0-9]+]], 31, v[[LO]]
 174 ; GCN-NOHSA: buffer_store_dwordx2 v[[[LO]]:[[HI]]]
 175 ; GCN-HSA: {{flat|global}}_store_dwordx2 v{{.+}}, v[[[LO]]:[[HI]]]
 176
 177
 178 ; EG: MEM_RAT
 179 ; EG: VTX_READ_32
 180 ; EG: ASHR {{[* ]*}}T{{[0-9]\.[XYZW]}}, T{{[0-9]\.[XYZW]}},  literal.
 181 ; EG: 31
 182 define amdgpu_kernel void @global_sextload_i32_to_i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 183   %ld = load i32, ptr addrspace(1) %in
 184   %ext = sext i32 %ld to i64
 185   store i64 %ext, ptr addrspace(1) %out
 186   ret void
 187 }
 188
 189 ; FUNC-LABEL: {{^}}global_zextload_v1i32_to_v1i64:
 190 ; GCN-NOHSA: buffer_load_dword
 191 ; GCN-NOHSA: buffer_store_dwordx2
 192
 193 ; GCN-HSA: {{flat|global}}_load_dword
 194 ; GCN-HSA: {{flat|global}}_store_dwordx2
 195 define amdgpu_kernel void @global_zextload_v1i32_to_v1i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 196   %ld = load <1 x i32>, ptr addrspace(1) %in
 197   %ext = zext <1 x i32> %ld to <1 x i64>
 198   store <1 x i64> %ext, ptr addrspace(1) %out
 199   ret void
 200 }
 201
 202 ; FUNC-LABEL: {{^}}global_sextload_v1i32_to_v1i64:
 203 ; GCN-NOHSA: buffer_load_dword v[[LO:[0-9]+]]
 204 ; GCN-HSA: {{flat|global}}_load_dword v[[LO:[0-9]+]]
 205 ; GCN: v_ashrrev_i32_e32 v[[HI:[0-9]+]], 31, v[[LO]]
 206 ; GCN-NOHSA: buffer_store_dwordx2 v[[[LO]]:[[HI]]]
 207 ; GCN-HSA: {{flat|global}}_store_dwordx2 v{{.+}}, v[[[LO]]:[[HI]]]
 208 define amdgpu_kernel void @global_sextload_v1i32_to_v1i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 209   %ld = load <1 x i32>, ptr addrspace(1) %in
 210   %ext = sext <1 x i32> %ld to <1 x i64>
 211   store <1 x i64> %ext, ptr addrspace(1) %out
 212   ret void
 213 }
 214
 215 ; FUNC-LABEL: {{^}}global_zextload_v2i32_to_v2i64:
 216 ; GCN-NOHSA: buffer_load_dwordx2
 217 ; GCN-NOHSA: buffer_store_dwordx4
 218
 219 ; GCN-HSA: {{flat|global}}_load_dwordx2
 220 ; GCN-HSA: {{flat|global}}_store_dwordx4
 221 define amdgpu_kernel void @global_zextload_v2i32_to_v2i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 222   %ld = load <2 x i32>, ptr addrspace(1) %in
 223   %ext = zext <2 x i32> %ld to <2 x i64>
 224   store <2 x i64> %ext, ptr addrspace(1) %out
 225   ret void
 226 }
 227
 228 ; FUNC-LABEL: {{^}}global_sextload_v2i32_to_v2i64:
 229 ; GCN-NOHSA: buffer_load_dwordx2
 230 ; GCN-HSA: {{flat|global}}_load_dwordx2
 231
 232 ; GCN-DAG: v_ashrrev_i32
 233 ; GCN-DAG: v_ashrrev_i32
 234
 235 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 236 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 237 define amdgpu_kernel void @global_sextload_v2i32_to_v2i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 238   %ld = load <2 x i32>, ptr addrspace(1) %in
 239   %ext = sext <2 x i32> %ld to <2 x i64>
 240   store <2 x i64> %ext, ptr addrspace(1) %out
 241   ret void
 242 }
 243
 244 ; FUNC-LABEL: {{^}}global_zextload_v4i32_to_v4i64:
 245 ; GCN-NOHSA: buffer_load_dwordx4
 246 ; GCN-NOHSA: buffer_store_dwordx4
 247 ; GCN-NOHSA: buffer_store_dwordx4
 248
 249 ; GCN-HSA: {{flat|global}}_load_dwordx4
 250 ; GCN-HSA: {{flat|global}}_store_dwordx4
 251 ; GCN-HSA: {{flat|global}}_store_dwordx4
 252 define amdgpu_kernel void @global_zextload_v4i32_to_v4i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 253   %ld = load <4 x i32>, ptr addrspace(1) %in
 254   %ext = zext <4 x i32> %ld to <4 x i64>
 255   store <4 x i64> %ext, ptr addrspace(1) %out
 256   ret void
 257 }
 258
 259 ; FUNC-LABEL: {{^}}global_sextload_v4i32_to_v4i64:
 260 ; GCN-NOHSA: buffer_load_dwordx4
 261 ; GCN-HSA: {{flat|global}}_load_dwordx4
 262
 263 ; GCN-DAG: v_ashrrev_i32
 264 ; GCN-DAG: v_ashrrev_i32
 265 ; GCN-DAG: v_ashrrev_i32
 266 ; GCN-DAG: v_ashrrev_i32
 267
 268 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 269 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 270
 271 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 272 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 273 define amdgpu_kernel void @global_sextload_v4i32_to_v4i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 274   %ld = load <4 x i32>, ptr addrspace(1) %in
 275   %ext = sext <4 x i32> %ld to <4 x i64>
 276   store <4 x i64> %ext, ptr addrspace(1) %out
 277   ret void
 278 }
 279
 280 ; FUNC-LABEL: {{^}}global_zextload_v8i32_to_v8i64:
 281 ; GCN-NOHSA: buffer_load_dwordx4
 282 ; GCN-NOHSA: buffer_load_dwordx4
 283
 284 ; GCN-HSA: {{flat|global}}_load_dwordx4
 285 ; GCN-HSA: {{flat|global}}_load_dwordx4
 286
 287 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 288 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 289 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 290 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 291
 292 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 293 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 294 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 295 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 296 define amdgpu_kernel void @global_zextload_v8i32_to_v8i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 297   %ld = load <8 x i32>, ptr addrspace(1) %in
 298   %ext = zext <8 x i32> %ld to <8 x i64>
 299   store <8 x i64> %ext, ptr addrspace(1) %out
 300   ret void
 301 }
 302
 303 ; FUNC-LABEL: {{^}}global_sextload_v8i32_to_v8i64:
 304 ; GCN-NOHSA: buffer_load_dwordx4
 305 ; GCN-NOHSA: buffer_load_dwordx4
 306
 307 ; GCN-HSA: {{flat|global}}_load_dwordx4
 308 ; GCN-HSA: {{flat|global}}_load_dwordx4
 309
 310 ; GCN-DAG: v_ashrrev_i32
 311 ; GCN-DAG: v_ashrrev_i32
 312 ; GCN-DAG: v_ashrrev_i32
 313 ; GCN-DAG: v_ashrrev_i32
 314 ; GCN-DAG: v_ashrrev_i32
 315 ; GCN-DAG: v_ashrrev_i32
 316 ; GCN-DAG: v_ashrrev_i32
 317 ; GCN-DAG: v_ashrrev_i32
 318
 319 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 320 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 321 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 322 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 323
 324 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 325 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 326 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 327 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 328 define amdgpu_kernel void @global_sextload_v8i32_to_v8i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 329   %ld = load <8 x i32>, ptr addrspace(1) %in
 330   %ext = sext <8 x i32> %ld to <8 x i64>
 331   store <8 x i64> %ext, ptr addrspace(1) %out
 332   ret void
 333 }
 334
 335 ; FUNC-LABEL: {{^}}global_sextload_v16i32_to_v16i64:
 336 ; GCN-NOHSA: buffer_load_dwordx4
 337 ; GCN-NOHSA: buffer_load_dwordx4
 338 ; GCN-NOHSA: buffer_load_dwordx4
 339 ; GCN-NOHSA: buffer_load_dwordx4
 340
 341 ; GCN-HSA: {{flat|global}}_load_dwordx4
 342 ; GCN-HSA: {{flat|global}}_load_dwordx4
 343 ; GCN-HSA: {{flat|global}}_load_dwordx4
 344 ; GCN-HSA: {{flat|global}}_load_dwordx4
 345
 346
 347 ; GCN-DAG: v_ashrrev_i32
 348 ; GCN-DAG: v_ashrrev_i32
 349 ; GCN-DAG: v_ashrrev_i32
 350 ; GCN-DAG: v_ashrrev_i32
 351 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 352 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 353
 354 ; GCN-DAG: v_ashrrev_i32
 355 ; GCN-DAG: v_ashrrev_i32
 356 ; GCN-DAG: v_ashrrev_i32
 357 ; GCN-DAG: v_ashrrev_i32
 358 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 359 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 360
 361 ; GCN-DAG: v_ashrrev_i32
 362 ; GCN-DAG: v_ashrrev_i32
 363 ; GCN-DAG: v_ashrrev_i32
 364 ; GCN-DAG: v_ashrrev_i32
 365 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 366 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 367
 368 ; GCN-DAG: v_ashrrev_i32
 369 ; GCN-DAG: v_ashrrev_i32
 370 ; GCN-DAG: v_ashrrev_i32
 371 ; GCN-DAG: v_ashrrev_i32
 372 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 373 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 374 define amdgpu_kernel void @global_sextload_v16i32_to_v16i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 375   %ld = load <16 x i32>, ptr addrspace(1) %in
 376   %ext = sext <16 x i32> %ld to <16 x i64>
 377   store <16 x i64> %ext, ptr addrspace(1) %out
 378   ret void
 379 }
 380
 381 ; FUNC-LABEL: {{^}}global_zextload_v16i32_to_v16i64
 382 ; GCN-NOHSA: buffer_load_dwordx4
 383 ; GCN-NOHSA: buffer_load_dwordx4
 384 ; GCN-NOHSA: buffer_load_dwordx4
 385 ; GCN-NOHSA: buffer_load_dwordx4
 386
 387 ; GCN-HSA: {{flat|global}}_load_dwordx4
 388 ; GCN-HSA: {{flat|global}}_load_dwordx4
 389 ; GCN-HSA: {{flat|global}}_load_dwordx4
 390 ; GCN-HSA: {{flat|global}}_load_dwordx4
 391
 392 ; GCN-NOHSA: buffer_store_dwordx4
 393 ; GCN-NOHSA: buffer_store_dwordx4
 394 ; GCN-NOHSA: buffer_store_dwordx4
 395 ; GCN-NOHSA: buffer_store_dwordx4
 396 ; GCN-NOHSA: buffer_store_dwordx4
 397 ; GCN-NOHSA: buffer_store_dwordx4
 398 ; GCN-NOHSA: buffer_store_dwordx4
 399 ; GCN-NOHSA: buffer_store_dwordx4
 400
 401 ; GCN-HSA: {{flat|global}}_store_dwordx4
 402 ; GCN-HSA: {{flat|global}}_store_dwordx4
 403 ; GCN-HSA: {{flat|global}}_store_dwordx4
 404 ; GCN-HSA: {{flat|global}}_store_dwordx4
 405 ; GCN-HSA: {{flat|global}}_store_dwordx4
 406 ; GCN-HSA: {{flat|global}}_store_dwordx4
 407 ; GCN-HSA: {{flat|global}}_store_dwordx4
 408 ; GCN-HSA: {{flat|global}}_store_dwordx4
 409 define amdgpu_kernel void @global_zextload_v16i32_to_v16i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 410   %ld = load <16 x i32>, ptr addrspace(1) %in
 411   %ext = zext <16 x i32> %ld to <16 x i64>
 412   store <16 x i64> %ext, ptr addrspace(1) %out
 413   ret void
 414 }
 415
 416 ; FUNC-LABEL: {{^}}global_sextload_v32i32_to_v32i64:
 417
 418 ; GCN-NOHSA: buffer_load_dwordx4
 419 ; GCN-NOHSA: buffer_load_dwordx4
 420 ; GCN-NOHSA: buffer_load_dwordx4
 421 ; GCN-NOHSA: buffer_load_dwordx4
 422 ; GCN-NOHSA: buffer_load_dwordx4
 423 ; GCN-NOHSA: buffer_load_dwordx4
 424 ; GCN-NOHSA: buffer_load_dwordx4
 425 ; GCN-NOHSA-DAG: buffer_load_dwordx4
 426
 427 ; GCN-HSA: {{flat|global}}_load_dwordx4
 428 ; GCN-HSA: {{flat|global}}_load_dwordx4
 429 ; GCN-HSA: {{flat|global}}_load_dwordx4
 430 ; GCN-HSA: {{flat|global}}_load_dwordx4
 431 ; GCN-HSA: {{flat|global}}_load_dwordx4
 432 ; GCN-HSA: {{flat|global}}_load_dwordx4
 433 ; GCN-HSA: {{flat|global}}_load_dwordx4
 434 ; GCN-HSA: {{flat|global}}_load_dwordx4
 435
 436 ; GCN-DAG: v_ashrrev_i32
 437 ; GCN-DAG: v_ashrrev_i32
 438 ; GCN-DAG: v_ashrrev_i32
 439 ; GCN-DAG: v_ashrrev_i32
 440 ; GCN-DAG: v_ashrrev_i32
 441 ; GCN-DAG: v_ashrrev_i32
 442 ; GCN-DAG: v_ashrrev_i32
 443 ; GCN-DAG: v_ashrrev_i32
 444 ; GCN-DAG: v_ashrrev_i32
 445 ; GCN-DAG: v_ashrrev_i32
 446 ; GCN-DAG: v_ashrrev_i32
 447 ; GCN-DAG: v_ashrrev_i32
 448 ; GCN-DAG: v_ashrrev_i32
 449 ; GCN-DAG: v_ashrrev_i32
 450 ; GCN-DAG: v_ashrrev_i32
 451 ; GCN-DAG: v_ashrrev_i32
 452 ; GCN-DAG: v_ashrrev_i32
 453 ; GCN-DAG: v_ashrrev_i32
 454 ; GCN-DAG: v_ashrrev_i32
 455 ; GCN-DAG: v_ashrrev_i32
 456 ; GCN-DAG: v_ashrrev_i32
 457 ; GCN-DAG: v_ashrrev_i32
 458 ; GCN-DAG: v_ashrrev_i32
 459 ; GCN-DAG: v_ashrrev_i32
 460 ; GCN-DAG: v_ashrrev_i32
 461 ; GCN-DAG: v_ashrrev_i32
 462 ; GCN-DAG: v_ashrrev_i32
 463 ; GCN-DAG: v_ashrrev_i32
 464 ; GCN-DAG: v_ashrrev_i32
 465 ; GCN-DAG: v_ashrrev_i32
 466 ; GCN-DAG: v_ashrrev_i32
 467 ; GCN-DAG: v_ashrrev_i32
 468
 469 ; GCN-NOHSA: buffer_store_dwordx4
 470 ; GCN-NOHSA: buffer_store_dwordx4
 471 ; GCN-NOHSA: buffer_store_dwordx4
 472 ; GCN-NOHSA: buffer_store_dwordx4
 473
 474 ; GCN-NOHSA: buffer_store_dwordx4
 475 ; GCN-NOHSA: buffer_store_dwordx4
 476 ; GCN-NOHSA: buffer_store_dwordx4
 477 ; GCN-NOHSA: buffer_store_dwordx4
 478
 479 ; GCN-NOHSA: buffer_store_dwordx4
 480 ; GCN-NOHSA: buffer_store_dwordx4
 481 ; GCN-NOHSA: buffer_store_dwordx4
 482 ; GCN-NOHSA: buffer_store_dwordx4
 483
 484 ; GCN-NOHSA: buffer_store_dwordx4
 485 ; GCN-NOHSA: buffer_store_dwordx4
 486 ; GCN-NOHSA: buffer_store_dwordx4
 487 ; GCN-NOHSA: buffer_store_dwordx4
 488
 489 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 490 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 491 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 492 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 493
 494 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 495 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 496 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 497 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 498
 499 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 500 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 501 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 502 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 503
 504 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 505 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 506 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 507 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 508
 509 define amdgpu_kernel void @global_sextload_v32i32_to_v32i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 510   %ld = load <32 x i32>, ptr addrspace(1) %in
 511   %ext = sext <32 x i32> %ld to <32 x i64>
 512   store <32 x i64> %ext, ptr addrspace(1) %out
 513   ret void
 514 }
 515
 516 ; FUNC-LABEL: {{^}}global_zextload_v32i32_to_v32i64:
 517 ; GCN-NOHSA: buffer_load_dwordx4
 518 ; GCN-NOHSA: buffer_load_dwordx4
 519 ; GCN-NOHSA: buffer_load_dwordx4
 520 ; GCN-NOHSA: buffer_load_dwordx4
 521 ; GCN-NOHSA: buffer_load_dwordx4
 522 ; GCN-NOHSA: buffer_load_dwordx4
 523 ; GCN-NOHSA: buffer_load_dwordx4
 524 ; GCN-NOHSA: buffer_load_dwordx4
 525
 526 ; GCN-HSA: {{flat|global}}_load_dwordx4
 527 ; GCN-HSA: {{flat|global}}_load_dwordx4
 528 ; GCN-HSA: {{flat|global}}_load_dwordx4
 529 ; GCN-HSA: {{flat|global}}_load_dwordx4
 530 ; GCN-HSA: {{flat|global}}_load_dwordx4
 531 ; GCN-HSA: {{flat|global}}_load_dwordx4
 532 ; GCN-HSA: {{flat|global}}_load_dwordx4
 533 ; GCN-HSA: {{flat|global}}_load_dwordx4
 534
 535
 536 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 537 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 538 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 539 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 540
 541 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 542 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 543 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 544 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 545
 546 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 547 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 548 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 549 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 550
 551 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 552 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 553 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 554 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 555
 556
 557 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 558 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 559 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 560 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 561
 562 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 563 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 564 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 565 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 566
 567 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 568 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 569 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 570 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 571
 572 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 573 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 574 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 575 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 576 define amdgpu_kernel void @global_zextload_v32i32_to_v32i64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 577   %ld = load <32 x i32>, ptr addrspace(1) %in
 578   %ext = zext <32 x i32> %ld to <32 x i64>
 579   store <32 x i64> %ext, ptr addrspace(1) %out
 580   ret void
 581 }
 582
 583 ; FUNC-LABEL: {{^}}global_load_v32i32:
 584 ; GCN-NOHSA: buffer_load_dwordx4
 585 ; GCN-NOHSA: buffer_load_dwordx4
 586 ; GCN-NOHSA: buffer_load_dwordx4
 587 ; GCN-NOHSA: buffer_load_dwordx4
 588 ; GCN-NOHSA: buffer_load_dwordx4
 589 ; GCN-NOHSA: buffer_load_dwordx4
 590 ; GCN-NOHSA: buffer_load_dwordx4
 591 ; GCN-NOHSA: buffer_load_dwordx4
 592
 593 ; GCN-HSA: {{flat|global}}_load_dwordx4
 594 ; GCN-HSA: {{flat|global}}_load_dwordx4
 595 ; GCN-HSA: {{flat|global}}_load_dwordx4
 596 ; GCN-HSA: {{flat|global}}_load_dwordx4
 597 ; GCN-HSA: {{flat|global}}_load_dwordx4
 598 ; GCN-HSA: {{flat|global}}_load_dwordx4
 599 ; GCN-HSA: {{flat|global}}_load_dwordx4
 600 ; GCN-HSA: {{flat|global}}_load_dwordx4
 601
 602
 603 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 604 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 605 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 606 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 607
 608 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 609 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 610 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 611 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 612
 613 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 614 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 615 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 616 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 617
 618 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 619 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 620 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 621 ; GCN-NOHSA-DAG: buffer_store_dwordx4
 622
 623 ; GCN-NOT: accvgpr
 624
 625 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 626 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 627 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 628 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 629
 630 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 631 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 632 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 633 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 634
 635 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 636 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 637 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 638 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 639
 640 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 641 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 642 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 643 ; GCN-HSA-DAG: {{flat|global}}_store_dwordx4
 644 define amdgpu_kernel void @global_load_v32i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
 645   %ld = load <32 x i32>, ptr addrspace(1) %in
 646   store <32 x i32> %ld, ptr addrspace(1) %out
 647   ret void
 648 }
 649
 650 attributes #0 = { nounwind }