test/CodeGen/AMDGPU/memory-legalizer-atomic-fence.ll

   1 ; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX6,GFX68 %s
   2 ; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX8,GFX68 %s
   3 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX8,GFX68 %s
   4 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+code-object-v3 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX10,GFX10WGP %s
   5 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+code-object-v3,+cumode -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX10,GFX10CU %s
   6
   7 ; FUNC-LABEL: {{^}}system_one_as_acquire:
   8 ; GCN:        %bb.0
   9 ; GCN-NOT:    ATOMIC_FENCE
  10 ; GFX6:       s_waitcnt vmcnt(0){{$}}
  11 ; GFX6-NEXT:  buffer_wbinvl1{{$}}
  12 ; GFX8:       s_waitcnt vmcnt(0){{$}}
  13 ; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
  14 ; GFX10:      s_waitcnt vmcnt(0){{$}}
  15 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
  16 ; GFX10-NEXT: buffer_gl0_inv{{$}}
  17 ; GFX10-NEXT: buffer_gl1_inv{{$}}
  18 ; GCN:        s_endpgm
  19 ; GFX10:         .amdhsa_kernel system_one_as_acquire
  20 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
  21 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
  22 ; GFX10-NOT:     .amdhsa_memory_ordered 0
  23 define amdgpu_kernel void @system_one_as_acquire() {
  24 entry:
  25   fence syncscope("one-as") acquire
  26   ret void
  27 }
  28
  29 ; FUNC-LABEL: {{^}}system_one_as_release:
  30 ; GCN:        %bb.0
  31 ; GCN-NOT:    ATOMIC_FENCE
  32 ; GCN:        s_waitcnt vmcnt(0){{$}}
  33 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
  34 ; GCN:        s_endpgm
  35 ; GFX10:         .amdhsa_kernel system_one_as_release
  36 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
  37 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
  38 ; GFX10-NOT:     .amdhsa_memory_ordered 0
  39 define amdgpu_kernel void @system_one_as_release() {
  40 entry:
  41   fence syncscope("one-as") release
  42   ret void
  43 }
  44
  45 ; FUNC-LABEL: {{^}}system_one_as_acq_rel:
  46 ; GCN:        %bb.0
  47 ; GCN-NOT:    ATOMIC_FENCE
  48 ; GCN:        s_waitcnt vmcnt(0){{$}}
  49 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
  50 ; GFX6:       buffer_wbinvl1{{$}}
  51 ; GFX8:       buffer_wbinvl1_vol{{$}}
  52 ; GFX10-NEXT: buffer_gl0_inv{{$}}
  53 ; GFX10-NEXT: buffer_gl1_inv{{$}}
  54 ; GCN:        s_endpgm
  55 ; GFX10:         .amdhsa_kernel system_one_as_acq_rel
  56 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
  57 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
  58 ; GFX10-NOT:     .amdhsa_memory_ordered 0
  59 define amdgpu_kernel void @system_one_as_acq_rel() {
  60 entry:
  61   fence syncscope("one-as") acq_rel
  62   ret void
  63 }
  64
  65 ; FUNC-LABEL: {{^}}system_one_as_seq_cst:
  66 ; GCN:        %bb.0
  67 ; GCN-NOT:    ATOMIC_FENCE
  68 ; GCN:        s_waitcnt vmcnt(0){{$}}
  69 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
  70 ; GFX6:       buffer_wbinvl1{{$}}
  71 ; GFX8:       buffer_wbinvl1_vol{{$}}
  72 ; GFX10-NEXT: buffer_gl0_inv{{$}}
  73 ; GFX10-NEXT: buffer_gl1_inv{{$}}
  74 ; GCN:        s_endpgm
  75 ; GFX10:         .amdhsa_kernel system_one_as_seq_cst
  76 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
  77 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
  78 ; GFX10-NOT:     .amdhsa_memory_ordered 0
  79 define amdgpu_kernel void @system_one_as_seq_cst() {
  80 entry:
  81   fence syncscope("one-as") seq_cst
  82   ret void
  83 }
  84
  85 ; FUNC-LABEL: {{^}}singlethread_one_as_acquire:
  86 ; GCN:        %bb.0
  87 ; GCN-NOT:    ATOMIC_FENCE
  88 ; GCN:        s_endpgm
  89 ; GFX10:         .amdhsa_kernel singlethread_one_as_acquire
  90 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
  91 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
  92 ; GFX10-NOT:     .amdhsa_memory_ordered 0
  93 define amdgpu_kernel void @singlethread_one_as_acquire() {
  94 entry:
  95   fence syncscope("singlethread-one-as") acquire
  96   ret void
  97 }
  98
  99 ; FUNC-LABEL: {{^}}singlethread_one_as_release:
 100 ; GCN:        %bb.0
 101 ; GCN-NOT:    ATOMIC_FENCE
 102 ; GCN:        s_endpgm
 103 ; GFX10:         .amdhsa_kernel singlethread_one_as_release
 104 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 105 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 106 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 107 define amdgpu_kernel void @singlethread_one_as_release() {
 108 entry:
 109   fence syncscope("singlethread-one-as") release
 110   ret void
 111 }
 112
 113 ; FUNC-LABEL: {{^}}singlethread_one_as_acq_rel:
 114 ; GCN:        %bb.0
 115 ; GCN-NOT:    ATOMIC_FENCE
 116 ; GCN:        s_endpgm
 117 ; GFX10:         .amdhsa_kernel singlethread_one_as_acq_rel
 118 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 119 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 120 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 121 define amdgpu_kernel void @singlethread_one_as_acq_rel() {
 122 entry:
 123   fence syncscope("singlethread-one-as") acq_rel
 124   ret void
 125 }
 126
 127 ; FUNC-LABEL: {{^}}singlethread_one_as_seq_cst:
 128 ; GCN:        %bb.0
 129 ; GCN-NOT:    ATOMIC_FENCE
 130 ; GCN:        s_endpgm
 131 ; GFX10:         .amdhsa_kernel singlethread_one_as_seq_cst
 132 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 133 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 134 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 135 define amdgpu_kernel void @singlethread_one_as_seq_cst() {
 136 entry:
 137   fence syncscope("singlethread-one-as") seq_cst
 138   ret void
 139 }
 140
 141 ; FUNC-LABEL: {{^}}agent_one_as_acquire:
 142 ; GCN:        %bb.0
 143 ; GCN-NOT:    ATOMIC_FENCE
 144 ; GFX6:       s_waitcnt vmcnt(0){{$}}
 145 ; GFX6-NEXT:  buffer_wbinvl1{{$}}
 146 ; GFX8:       s_waitcnt vmcnt(0){{$}}
 147 ; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
 148 ; GFX10:      s_waitcnt vmcnt(0){{$}}
 149 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 150 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 151 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 152 ; GCN:        s_endpgm
 153 ; GFX10:         .amdhsa_kernel agent_one_as_acquire
 154 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 155 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 156 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 157 define amdgpu_kernel void @agent_one_as_acquire() {
 158 entry:
 159   fence syncscope("agent-one-as") acquire
 160   ret void
 161 }
 162
 163 ; FUNC-LABEL: {{^}}agent_one_as_release:
 164 ; GCN:        %bb.0
 165 ; GCN-NOT:    ATOMIC_FENCE
 166 ; GCN:        s_waitcnt vmcnt(0){{$}}
 167 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 168 ; GCN:        s_endpgm
 169 ; GFX10:         .amdhsa_kernel agent_one_as_release
 170 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 171 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 172 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 173 define amdgpu_kernel void @agent_one_as_release() {
 174 entry:
 175   fence syncscope("agent-one-as") release
 176   ret void
 177 }
 178
 179 ; FUNC-LABEL: {{^}}agent_one_as_acq_rel:
 180 ; GCN:        %bb.0
 181 ; GCN-NOT:    ATOMIC_FENCE
 182 ; GCN:        s_waitcnt vmcnt(0){{$}}
 183 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 184 ; GFX6:       buffer_wbinvl1{{$}}
 185 ; GFX8:       buffer_wbinvl1_vol{{$}}
 186 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 187 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 188 ; GCN:        s_endpgm
 189 ; GFX10:         .amdhsa_kernel agent_one_as_acq_rel
 190 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 191 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 192 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 193 define amdgpu_kernel void @agent_one_as_acq_rel() {
 194 entry:
 195   fence syncscope("agent-one-as") acq_rel
 196   ret void
 197 }
 198
 199 ; FUNC-LABEL: {{^}}agent_one_as_seq_cst:
 200 ; GCN:        %bb.0
 201 ; GCN-NOT:    ATOMIC_FENCE
 202 ; GCN:        s_waitcnt vmcnt(0){{$}}
 203 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 204 ; GFX6:       buffer_wbinvl1{{$}}
 205 ; GFX8:       buffer_wbinvl1_vol{{$}}
 206 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 207 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 208 ; GCN:        s_endpgm
 209 ; GFX10:         .amdhsa_kernel agent_one_as_seq_cst
 210 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 211 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 212 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 213 define amdgpu_kernel void @agent_one_as_seq_cst() {
 214 entry:
 215   fence syncscope("agent-one-as") seq_cst
 216   ret void
 217 }
 218
 219 ; FUNC-LABEL:    {{^}}workgroup_one_as_acquire:
 220 ; GCN:           %bb.0
 221 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 222 ; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
 223 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 224 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 225 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 226 ; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
 227 ; GFX10CU-NOT:   buffer_gl0_inv{{$}}
 228 ; GCN-NOT:       ATOMIC_FENCE
 229 ; GCN:           s_endpgm
 230 ; GFX10:         .amdhsa_kernel workgroup_one_as_acquire
 231 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 232 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 233 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 234 define amdgpu_kernel void @workgroup_one_as_acquire() {
 235 entry:
 236   fence syncscope("workgroup-one-as") acquire
 237   ret void
 238 }
 239
 240 ; FUNC-LABEL:    {{^}}workgroup_one_as_release:
 241 ; GCN:           %bb.0
 242 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 243 ; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
 244 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 245 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 246 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 247 ; GFX10-NOT:     buffer_gl0_inv
 248 ; GCN-NOT:       ATOMIC_FENCE
 249 ; GCN:           s_endpgm
 250 ; GFX10:         .amdhsa_kernel workgroup_one_as_release
 251 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 252 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 253 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 254 define amdgpu_kernel void @workgroup_one_as_release() {
 255 entry:
 256   fence syncscope("workgroup-one-as") release
 257   ret void
 258 }
 259
 260 ; FUNC-LABEL:    {{^}}workgroup_one_as_acq_rel:
 261 ; GCN:           %bb.0
 262 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 263 ; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
 264 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 265 ; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
 266 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 267 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 268 ; GFX10CU-NOT:   buffer_gl0_inv{{$}}
 269 ; GCN-NOT:       ATOMIC_FENCE
 270 ; GCN:           s_endpgm
 271 ; GFX10:         .amdhsa_kernel workgroup_one_as_acq_rel
 272 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 273 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 274 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 275 define amdgpu_kernel void @workgroup_one_as_acq_rel() {
 276 entry:
 277   fence syncscope("workgroup-one-as") acq_rel
 278   ret void
 279 }
 280
 281 ; FUNC-LABEL:    {{^}}workgroup_one_as_seq_cst:
 282 ; GCN:           %bb.0
 283 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 284 ; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
 285 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 286 ; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
 287 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 288 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 289 ; GFX10CU-NOT:   buffer_gl0_inv{{$}}
 290 ; GCN-NOT:       ATOMIC_FENCE
 291 ; GCN:           s_endpgm
 292 ; GFX10:         .amdhsa_kernel workgroup_one_as_seq_cst
 293 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 294 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 295 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 296 define amdgpu_kernel void @workgroup_one_as_seq_cst() {
 297 entry:
 298   fence syncscope("workgroup-one-as") seq_cst
 299   ret void
 300 }
 301
 302 ; FUNC-LABEL: {{^}}wavefront_one_as_acquire:
 303 ; GCN:        %bb.0
 304 ; GCN-NOT:    ATOMIC_FENCE
 305 ; GCN:        s_endpgm
 306 ; GFX10:         .amdhsa_kernel wavefront_one_as_acquire
 307 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 308 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 309 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 310 define amdgpu_kernel void @wavefront_one_as_acquire() {
 311 entry:
 312   fence syncscope("wavefront-one-as") acquire
 313   ret void
 314 }
 315
 316 ; FUNC-LABEL: {{^}}wavefront_one_as_release:
 317 ; GCN:        %bb.0
 318 ; GCN-NOT:    ATOMIC_FENCE
 319 ; GCN:        s_endpgm
 320 ; GFX10:         .amdhsa_kernel wavefront_one_as_release
 321 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 322 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 323 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 324 define amdgpu_kernel void @wavefront_one_as_release() {
 325 entry:
 326   fence syncscope("wavefront-one-as") release
 327   ret void
 328 }
 329
 330 ; FUNC-LABEL: {{^}}wavefront_one_as_acq_rel:
 331 ; GCN:        %bb.0
 332 ; GCN-NOT:    ATOMIC_FENCE
 333 ; GCN:        s_endpgm
 334 ; GFX10:         .amdhsa_kernel wavefront_one_as_acq_rel
 335 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 336 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 337 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 338 define amdgpu_kernel void @wavefront_one_as_acq_rel() {
 339 entry:
 340   fence syncscope("wavefront-one-as") acq_rel
 341   ret void
 342 }
 343
 344 ; FUNC-LABEL: {{^}}wavefront_one_as_seq_cst:
 345 ; GCN:        %bb.0
 346 ; GCN-NOT:    ATOMIC_FENCE
 347 ; GCN:        s_endpgm
 348 ; GFX10:         .amdhsa_kernel wavefront_one_as_seq_cst
 349 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 350 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 351 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 352 define amdgpu_kernel void @wavefront_one_as_seq_cst() {
 353 entry:
 354   fence syncscope("wavefront-one-as") seq_cst
 355   ret void
 356 }
 357
 358 ; FUNC-LABEL: {{^}}system_acquire:
 359 ; GCN:        %bb.0
 360 ; GCN-NOT:    ATOMIC_FENCE
 361 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 362 ; GFX6-NEXT:  buffer_wbinvl1{{$}}
 363 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 364 ; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
 365 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 366 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 367 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 368 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 369 ; GCN:        s_endpgm
 370 ; GFX10:         .amdhsa_kernel system_acquire
 371 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 372 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 373 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 374 define amdgpu_kernel void @system_acquire() {
 375 entry:
 376   fence acquire
 377   ret void
 378 }
 379
 380 ; FUNC-LABEL: {{^}}system_release:
 381 ; GCN:        %bb.0
 382 ; GCN-NOT:    ATOMIC_FENCE
 383 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 384 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 385 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 386 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 387 ; GCN:        s_endpgm
 388 ; GFX10:         .amdhsa_kernel system_release
 389 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 390 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 391 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 392 define amdgpu_kernel void @system_release() {
 393 entry:
 394   fence release
 395   ret void
 396 }
 397
 398 ; FUNC-LABEL: {{^}}system_acq_rel:
 399 ; GCN:        %bb.0
 400 ; GCN-NOT:    ATOMIC_FENCE
 401 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 402 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 403 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 404 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 405 ; GFX6:       buffer_wbinvl1{{$}}
 406 ; GFX8:       buffer_wbinvl1_vol{{$}}
 407 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 408 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 409 ; GCN:        s_endpgm
 410 ; GFX10:         .amdhsa_kernel system_acq_rel
 411 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 412 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 413 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 414 define amdgpu_kernel void @system_acq_rel() {
 415 entry:
 416   fence acq_rel
 417   ret void
 418 }
 419
 420 ; FUNC-LABEL: {{^}}system_seq_cst:
 421 ; GCN:        %bb.0
 422 ; GCN-NOT:    ATOMIC_FENCE
 423 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 424 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 425 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 426 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 427 ; GFX6:       buffer_wbinvl1{{$}}
 428 ; GFX8:       buffer_wbinvl1_vol{{$}}
 429 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 430 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 431 ; GCN:        s_endpgm
 432 ; GFX10:         .amdhsa_kernel system_seq_cst
 433 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 434 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 435 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 436 define amdgpu_kernel void @system_seq_cst() {
 437 entry:
 438   fence seq_cst
 439   ret void
 440 }
 441
 442 ; FUNC-LABEL: {{^}}singlethread_acquire:
 443 ; GCN:        %bb.0
 444 ; GCN-NOT:    ATOMIC_FENCE
 445 ; GCN:        s_endpgm
 446 ; GFX10:         .amdhsa_kernel singlethread_acquire
 447 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 448 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 449 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 450 define amdgpu_kernel void @singlethread_acquire() {
 451 entry:
 452   fence syncscope("singlethread") acquire
 453   ret void
 454 }
 455
 456 ; FUNC-LABEL: {{^}}singlethread_release:
 457 ; GCN:        %bb.0
 458 ; GCN-NOT:    ATOMIC_FENCE
 459 ; GCN:        s_endpgm
 460 ; GFX10:         .amdhsa_kernel singlethread_release
 461 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 462 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 463 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 464 define amdgpu_kernel void @singlethread_release() {
 465 entry:
 466   fence syncscope("singlethread") release
 467   ret void
 468 }
 469
 470 ; FUNC-LABEL: {{^}}singlethread_acq_rel:
 471 ; GCN:        %bb.0
 472 ; GCN-NOT:    ATOMIC_FENCE
 473 ; GCN:        s_endpgm
 474 ; GFX10:         .amdhsa_kernel singlethread_acq_rel
 475 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 476 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 477 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 478 define amdgpu_kernel void @singlethread_acq_rel() {
 479 entry:
 480   fence syncscope("singlethread") acq_rel
 481   ret void
 482 }
 483
 484 ; FUNC-LABEL: {{^}}singlethread_seq_cst:
 485 ; GCN:        %bb.0
 486 ; GCN-NOT:    ATOMIC_FENCE
 487 ; GCN:        s_endpgm
 488 ; GFX10:         .amdhsa_kernel singlethread_seq_cst
 489 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 490 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 491 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 492 define amdgpu_kernel void @singlethread_seq_cst() {
 493 entry:
 494   fence syncscope("singlethread") seq_cst
 495   ret void
 496 }
 497
 498 ; FUNC-LABEL: {{^}}agent_acquire:
 499 ; GCN:        %bb.0
 500 ; GCN-NOT:    ATOMIC_FENCE
 501 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 502 ; GFX6-NEXT:  buffer_wbinvl1{{$}}
 503 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 504 ; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
 505 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 506 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 507 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 508 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 509 ; GCN:        s_endpgm
 510 ; GFX10:         .amdhsa_kernel agent_acquire
 511 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 512 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 513 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 514 define amdgpu_kernel void @agent_acquire() {
 515 entry:
 516   fence syncscope("agent") acquire
 517   ret void
 518 }
 519
 520 ; FUNC-LABEL: {{^}}agent_release:
 521 ; GCN:        %bb.0
 522 ; GCN-NOT:    ATOMIC_FENCE
 523 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 524 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 525 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 526 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 527 ; GCN:        s_endpgm
 528 ; GFX10:         .amdhsa_kernel agent_release
 529 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 530 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 531 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 532 define amdgpu_kernel void @agent_release() {
 533 entry:
 534   fence syncscope("agent") release
 535   ret void
 536 }
 537
 538 ; FUNC-LABEL: {{^}}agent_acq_rel:
 539 ; GCN:        %bb.0
 540 ; GCN-NOT:    ATOMIC_FENCE
 541 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 542 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 543 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 544 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 545 ; GFX6:       buffer_wbinvl1{{$}}
 546 ; GFX8:       buffer_wbinvl1_vol{{$}}
 547 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 548 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 549 ; GCN:        s_endpgm
 550 ; GFX10:         .amdhsa_kernel agent_acq_rel
 551 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 552 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 553 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 554 define amdgpu_kernel void @agent_acq_rel() {
 555 entry:
 556   fence syncscope("agent") acq_rel
 557   ret void
 558 }
 559
 560 ; FUNC-LABEL: {{^}}agent_seq_cst:
 561 ; GCN:        %bb.0
 562 ; GCN-NOT:    ATOMIC_FENCE
 563 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 564 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 565 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 566 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 567 ; GFX6:       buffer_wbinvl1{{$}}
 568 ; GFX8:       buffer_wbinvl1_vol{{$}}
 569 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 570 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 571 ; GCN:        s_endpgm
 572 ; GFX10:         .amdhsa_kernel agent_seq_cst
 573 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 574 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 575 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 576 define amdgpu_kernel void @agent_seq_cst() {
 577 entry:
 578   fence syncscope("agent") seq_cst
 579   ret void
 580 }
 581
 582 ; FUNC-LABEL:    {{^}}workgroup_acquire:
 583 ; GCN:           %bb.0
 584 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 585 ; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 586 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 587 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 588 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 589 ; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
 590 ; GFX10CU-NOT:   buffer_gl0_inv{{$}}
 591 ; GCN-NOT:       ATOMIC_FENCE
 592 ; GCN:           s_endpgm
 593 ; GFX10:         .amdhsa_kernel workgroup_acquire
 594 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 595 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 596 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 597 define amdgpu_kernel void @workgroup_acquire() {
 598 entry:
 599   fence syncscope("workgroup") acquire
 600   ret void
 601 }
 602
 603 ; FUNC-LABEL:    {{^}}workgroup_release:
 604 ; GCN:           %bb.0
 605 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 606 ; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 607 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 608 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 609 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 610 ; GFX10-NOT:     buffer_gl0_inv
 611 ; GCN-NOT:       ATOMIC_FENCE
 612 ; GCN:           s_endpgm
 613 ; GFX10:         .amdhsa_kernel workgroup_release
 614 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 615 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 616 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 617 define amdgpu_kernel void @workgroup_release() {
 618 entry:
 619   fence syncscope("workgroup") release
 620   ret void
 621 }
 622
 623 ; FUNC-LABEL:    {{^}}workgroup_acq_rel:
 624 ; GCN:           %bb.0
 625 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 626 ; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 627 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 628 ; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
 629 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 630 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 631 ; GFX10CU-NOT:   buffer_gl0_inv{{$}}
 632 ; GCN-NOT:       ATOMIC_FENCE
 633 ; GCN:           s_endpgm
 634 ; GFX10:         .amdhsa_kernel workgroup_acq_rel
 635 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 636 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 637 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 638 define amdgpu_kernel void @workgroup_acq_rel() {
 639 entry:
 640   fence syncscope("workgroup") acq_rel
 641   ret void
 642 }
 643
 644 ; FUNC-LABEL:    {{^}}workgroup_seq_cst:
 645 ; GCN:           %bb.0
 646 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 647 ; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 648 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 649 ; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
 650 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 651 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 652 ; GFX10CU-NOT:   buffer_gl0_inv{{$}}
 653 ; GCN-NOT:       ATOMIC_FENCE
 654 ; GCN:           s_endpgm
 655 ; GFX10:         .amdhsa_kernel workgroup_seq_cst
 656 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 657 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 658 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 659 define amdgpu_kernel void @workgroup_seq_cst() {
 660 entry:
 661   fence syncscope("workgroup") seq_cst
 662   ret void
 663 }
 664
 665 ; FUNC-LABEL: {{^}}wavefront_acquire:
 666 ; GCN:        %bb.0
 667 ; GCN-NOT:    ATOMIC_FENCE
 668 ; GCN:        s_endpgm
 669 ; GFX10:         .amdhsa_kernel wavefront_acquire
 670 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 671 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 672 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 673 define amdgpu_kernel void @wavefront_acquire() {
 674 entry:
 675   fence syncscope("wavefront") acquire
 676   ret void
 677 }
 678
 679 ; FUNC-LABEL: {{^}}wavefront_release:
 680 ; GCN:        %bb.0
 681 ; GCN-NOT:    ATOMIC_FENCE
 682 ; GCN:        s_endpgm
 683 ; GFX10:         .amdhsa_kernel wavefront_release
 684 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 685 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 686 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 687 define amdgpu_kernel void @wavefront_release() {
 688 entry:
 689   fence syncscope("wavefront") release
 690   ret void
 691 }
 692
 693 ; FUNC-LABEL: {{^}}wavefront_acq_rel:
 694 ; GCN:        %bb.0
 695 ; GCN-NOT:    ATOMIC_FENCE
 696 ; GCN:        s_endpgm
 697 ; GFX10:         .amdhsa_kernel wavefront_acq_rel
 698 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 699 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 700 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 701 define amdgpu_kernel void @wavefront_acq_rel() {
 702 entry:
 703   fence syncscope("wavefront") acq_rel
 704   ret void
 705 }
 706
 707 ; FUNC-LABEL: {{^}}wavefront_seq_cst:
 708 ; GCN:        %bb.0
 709 ; GCN-NOT:    ATOMIC_FENCE
 710 ; GCN:        s_endpgm
 711 ; GFX10:         .amdhsa_kernel wavefront_seq_cst
 712 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 713 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 714 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 715 define amdgpu_kernel void @wavefront_seq_cst() {
 716 entry:
 717   fence syncscope("wavefront") seq_cst
 718   ret void
 719 }