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

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