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

   1 ; RUN: llc -global-isel -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX6,GFX68 %s
   2 ; RUN: llc -global-isel -mtriple=amdgcn-amd- -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX8,GFX68 %s
   3 ; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX8,GFX68 %s
   4 ; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX10,GFX10WGP %s
   5 ; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX10,GFX10CU %s
   6 ; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX10,GFX10WGP %s
   7 ; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -mattr=+cumode -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX10,GFX10CU %s
   8
   9 ; FUNC-LABEL: {{^}}system_one_as_acquire:
  10 ; GCN:        %bb.0
  11 ; GCN-NOT:    ATOMIC_FENCE
  12 ; GFX6:       s_waitcnt vmcnt(0){{$}}
  13 ; GFX6-NEXT:  buffer_wbinvl1{{$}}
  14 ; GFX8:       s_waitcnt vmcnt(0){{$}}
  15 ; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
  16 ; GFX10:      s_waitcnt vmcnt(0){{$}}
  17 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
  18 ; GFX10-NEXT: buffer_gl0_inv{{$}}
  19 ; GFX10-NEXT: buffer_gl1_inv{{$}}
  20 ; GCN:        s_endpgm
  21 ; GFX10:         .amdhsa_kernel system_one_as_acquire
  22 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
  23 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
  24 ; GFX10-NOT:     .amdhsa_memory_ordered 0
  25 define amdgpu_kernel void @system_one_as_acquire() {
  26 entry:
  27   fence syncscope("one-as") acquire
  28   ret void
  29 }
  30
  31 ; FUNC-LABEL: {{^}}system_one_as_release:
  32 ; GCN:        %bb.0
  33 ; GCN-NOT:    ATOMIC_FENCE
  34 ; GCN:        s_waitcnt vmcnt(0){{$}}
  35 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
  36 ; GCN:        s_endpgm
  37 ; GFX10:         .amdhsa_kernel system_one_as_release
  38 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
  39 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
  40 ; GFX10-NOT:     .amdhsa_memory_ordered 0
  41 define amdgpu_kernel void @system_one_as_release() {
  42 entry:
  43   fence syncscope("one-as") release
  44   ret void
  45 }
  46
  47 ; FUNC-LABEL: {{^}}system_one_as_acq_rel:
  48 ; GCN:        %bb.0
  49 ; GCN-NOT:    ATOMIC_FENCE
  50 ; GCN:        s_waitcnt vmcnt(0){{$}}
  51 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
  52 ; GFX6:       buffer_wbinvl1{{$}}
  53 ; GFX8:       buffer_wbinvl1_vol{{$}}
  54 ; GFX10-NEXT: buffer_gl0_inv{{$}}
  55 ; GFX10-NEXT: buffer_gl1_inv{{$}}
  56 ; GCN:        s_endpgm
  57 ; GFX10:         .amdhsa_kernel system_one_as_acq_rel
  58 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
  59 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
  60 ; GFX10-NOT:     .amdhsa_memory_ordered 0
  61 define amdgpu_kernel void @system_one_as_acq_rel() {
  62 entry:
  63   fence syncscope("one-as") acq_rel
  64   ret void
  65 }
  66
  67 ; FUNC-LABEL: {{^}}system_one_as_seq_cst:
  68 ; GCN:        %bb.0
  69 ; GCN-NOT:    ATOMIC_FENCE
  70 ; GCN:        s_waitcnt vmcnt(0){{$}}
  71 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
  72 ; GFX6:       buffer_wbinvl1{{$}}
  73 ; GFX8:       buffer_wbinvl1_vol{{$}}
  74 ; GFX10-NEXT: buffer_gl0_inv{{$}}
  75 ; GFX10-NEXT: buffer_gl1_inv{{$}}
  76 ; GCN:        s_endpgm
  77 ; GFX10:         .amdhsa_kernel system_one_as_seq_cst
  78 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
  79 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
  80 ; GFX10-NOT:     .amdhsa_memory_ordered 0
  81 define amdgpu_kernel void @system_one_as_seq_cst() {
  82 entry:
  83   fence syncscope("one-as") seq_cst
  84   ret void
  85 }
  86
  87 ; FUNC-LABEL: {{^}}singlethread_one_as_acquire:
  88 ; GCN:        %bb.0
  89 ; GCN-NOT:    ATOMIC_FENCE
  90 ; GCN:        s_endpgm
  91 ; GFX10:         .amdhsa_kernel singlethread_one_as_acquire
  92 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
  93 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
  94 ; GFX10-NOT:     .amdhsa_memory_ordered 0
  95 define amdgpu_kernel void @singlethread_one_as_acquire() {
  96 entry:
  97   fence syncscope("singlethread-one-as") acquire
  98   ret void
  99 }
 100
 101 ; FUNC-LABEL: {{^}}singlethread_one_as_release:
 102 ; GCN:        %bb.0
 103 ; GCN-NOT:    ATOMIC_FENCE
 104 ; GCN:        s_endpgm
 105 ; GFX10:         .amdhsa_kernel singlethread_one_as_release
 106 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 107 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 108 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 109 define amdgpu_kernel void @singlethread_one_as_release() {
 110 entry:
 111   fence syncscope("singlethread-one-as") release
 112   ret void
 113 }
 114
 115 ; FUNC-LABEL: {{^}}singlethread_one_as_acq_rel:
 116 ; GCN:        %bb.0
 117 ; GCN-NOT:    ATOMIC_FENCE
 118 ; GCN:        s_endpgm
 119 ; GFX10:         .amdhsa_kernel singlethread_one_as_acq_rel
 120 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 121 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 122 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 123 define amdgpu_kernel void @singlethread_one_as_acq_rel() {
 124 entry:
 125   fence syncscope("singlethread-one-as") acq_rel
 126   ret void
 127 }
 128
 129 ; FUNC-LABEL: {{^}}singlethread_one_as_seq_cst:
 130 ; GCN:        %bb.0
 131 ; GCN-NOT:    ATOMIC_FENCE
 132 ; GCN:        s_endpgm
 133 ; GFX10:         .amdhsa_kernel singlethread_one_as_seq_cst
 134 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 135 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 136 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 137 define amdgpu_kernel void @singlethread_one_as_seq_cst() {
 138 entry:
 139   fence syncscope("singlethread-one-as") seq_cst
 140   ret void
 141 }
 142
 143 ; FUNC-LABEL: {{^}}agent_one_as_acquire:
 144 ; GCN:        %bb.0
 145 ; GCN-NOT:    ATOMIC_FENCE
 146 ; GFX6:       s_waitcnt vmcnt(0){{$}}
 147 ; GFX6-NEXT:  buffer_wbinvl1{{$}}
 148 ; GFX8:       s_waitcnt vmcnt(0){{$}}
 149 ; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
 150 ; GFX10:      s_waitcnt vmcnt(0){{$}}
 151 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 152 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 153 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 154 ; GCN:        s_endpgm
 155 ; GFX10:         .amdhsa_kernel agent_one_as_acquire
 156 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 157 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 158 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 159 define amdgpu_kernel void @agent_one_as_acquire() {
 160 entry:
 161   fence syncscope("agent-one-as") acquire
 162   ret void
 163 }
 164
 165 ; FUNC-LABEL: {{^}}agent_one_as_release:
 166 ; GCN:        %bb.0
 167 ; GCN-NOT:    ATOMIC_FENCE
 168 ; GCN:        s_waitcnt vmcnt(0){{$}}
 169 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 170 ; GCN:        s_endpgm
 171 ; GFX10:         .amdhsa_kernel agent_one_as_release
 172 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 173 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 174 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 175 define amdgpu_kernel void @agent_one_as_release() {
 176 entry:
 177   fence syncscope("agent-one-as") release
 178   ret void
 179 }
 180
 181 ; FUNC-LABEL: {{^}}agent_one_as_acq_rel:
 182 ; GCN:        %bb.0
 183 ; GCN-NOT:    ATOMIC_FENCE
 184 ; GCN:        s_waitcnt vmcnt(0){{$}}
 185 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 186 ; GFX6:       buffer_wbinvl1{{$}}
 187 ; GFX8:       buffer_wbinvl1_vol{{$}}
 188 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 189 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 190 ; GCN:        s_endpgm
 191 ; GFX10:         .amdhsa_kernel agent_one_as_acq_rel
 192 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 193 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 194 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 195 define amdgpu_kernel void @agent_one_as_acq_rel() {
 196 entry:
 197   fence syncscope("agent-one-as") acq_rel
 198   ret void
 199 }
 200
 201 ; FUNC-LABEL: {{^}}agent_one_as_seq_cst:
 202 ; GCN:        %bb.0
 203 ; GCN-NOT:    ATOMIC_FENCE
 204 ; GCN:        s_waitcnt vmcnt(0){{$}}
 205 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 206 ; GFX6:       buffer_wbinvl1{{$}}
 207 ; GFX8:       buffer_wbinvl1_vol{{$}}
 208 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 209 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 210 ; GCN:        s_endpgm
 211 ; GFX10:         .amdhsa_kernel agent_one_as_seq_cst
 212 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 213 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 214 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 215 define amdgpu_kernel void @agent_one_as_seq_cst() {
 216 entry:
 217   fence syncscope("agent-one-as") seq_cst
 218   ret void
 219 }
 220
 221 ; FUNC-LABEL:    {{^}}workgroup_one_as_acquire:
 222 ; GCN:           %bb.0
 223 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 224 ; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
 225 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 226 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 227 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 228 ; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
 229 ; GFX10CU-NOT:   buffer_gl0_inv{{$}}
 230 ; GCN-NOT:       ATOMIC_FENCE
 231 ; GCN:           s_endpgm
 232 ; GFX10:         .amdhsa_kernel workgroup_one_as_acquire
 233 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 234 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 235 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 236 define amdgpu_kernel void @workgroup_one_as_acquire() {
 237 entry:
 238   fence syncscope("workgroup-one-as") acquire
 239   ret void
 240 }
 241
 242 ; FUNC-LABEL:    {{^}}workgroup_one_as_release:
 243 ; GCN:           %bb.0
 244 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 245 ; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
 246 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 247 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 248 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 249 ; GFX10-NOT:     buffer_gl0_inv
 250 ; GCN-NOT:       ATOMIC_FENCE
 251 ; GCN:           s_endpgm
 252 ; GFX10:         .amdhsa_kernel workgroup_one_as_release
 253 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 254 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 255 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 256 define amdgpu_kernel void @workgroup_one_as_release() {
 257 entry:
 258   fence syncscope("workgroup-one-as") release
 259   ret void
 260 }
 261
 262 ; FUNC-LABEL:    {{^}}workgroup_one_as_acq_rel:
 263 ; GCN:           %bb.0
 264 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 265 ; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
 266 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 267 ; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
 268 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 269 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 270 ; GFX10CU-NOT:   buffer_gl0_inv{{$}}
 271 ; GCN-NOT:       ATOMIC_FENCE
 272 ; GCN:           s_endpgm
 273 ; GFX10:         .amdhsa_kernel workgroup_one_as_acq_rel
 274 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 275 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 276 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 277 define amdgpu_kernel void @workgroup_one_as_acq_rel() {
 278 entry:
 279   fence syncscope("workgroup-one-as") acq_rel
 280   ret void
 281 }
 282
 283 ; FUNC-LABEL:    {{^}}workgroup_one_as_seq_cst:
 284 ; GCN:           %bb.0
 285 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 286 ; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
 287 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 288 ; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
 289 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 290 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 291 ; GFX10CU-NOT:   buffer_gl0_inv{{$}}
 292 ; GCN-NOT:       ATOMIC_FENCE
 293 ; GCN:           s_endpgm
 294 ; GFX10:         .amdhsa_kernel workgroup_one_as_seq_cst
 295 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 296 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 297 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 298 define amdgpu_kernel void @workgroup_one_as_seq_cst() {
 299 entry:
 300   fence syncscope("workgroup-one-as") seq_cst
 301   ret void
 302 }
 303
 304 ; FUNC-LABEL: {{^}}wavefront_one_as_acquire:
 305 ; GCN:        %bb.0
 306 ; GCN-NOT:    ATOMIC_FENCE
 307 ; GCN:        s_endpgm
 308 ; GFX10:         .amdhsa_kernel wavefront_one_as_acquire
 309 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 310 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 311 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 312 define amdgpu_kernel void @wavefront_one_as_acquire() {
 313 entry:
 314   fence syncscope("wavefront-one-as") acquire
 315   ret void
 316 }
 317
 318 ; FUNC-LABEL: {{^}}wavefront_one_as_release:
 319 ; GCN:        %bb.0
 320 ; GCN-NOT:    ATOMIC_FENCE
 321 ; GCN:        s_endpgm
 322 ; GFX10:         .amdhsa_kernel wavefront_one_as_release
 323 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 324 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 325 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 326 define amdgpu_kernel void @wavefront_one_as_release() {
 327 entry:
 328   fence syncscope("wavefront-one-as") release
 329   ret void
 330 }
 331
 332 ; FUNC-LABEL: {{^}}wavefront_one_as_acq_rel:
 333 ; GCN:        %bb.0
 334 ; GCN-NOT:    ATOMIC_FENCE
 335 ; GCN:        s_endpgm
 336 ; GFX10:         .amdhsa_kernel wavefront_one_as_acq_rel
 337 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 338 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 339 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 340 define amdgpu_kernel void @wavefront_one_as_acq_rel() {
 341 entry:
 342   fence syncscope("wavefront-one-as") acq_rel
 343   ret void
 344 }
 345
 346 ; FUNC-LABEL: {{^}}wavefront_one_as_seq_cst:
 347 ; GCN:        %bb.0
 348 ; GCN-NOT:    ATOMIC_FENCE
 349 ; GCN:        s_endpgm
 350 ; GFX10:         .amdhsa_kernel wavefront_one_as_seq_cst
 351 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 352 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 353 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 354 define amdgpu_kernel void @wavefront_one_as_seq_cst() {
 355 entry:
 356   fence syncscope("wavefront-one-as") seq_cst
 357   ret void
 358 }
 359
 360 ; FUNC-LABEL: {{^}}system_acquire:
 361 ; GCN:        %bb.0
 362 ; GCN-NOT:    ATOMIC_FENCE
 363 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 364 ; GFX6-NEXT:  buffer_wbinvl1{{$}}
 365 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 366 ; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
 367 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 368 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 369 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 370 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 371 ; GCN:        s_endpgm
 372 ; GFX10:         .amdhsa_kernel system_acquire
 373 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 374 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 375 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 376 define amdgpu_kernel void @system_acquire() {
 377 entry:
 378   fence acquire
 379   ret void
 380 }
 381
 382 ; FUNC-LABEL: {{^}}system_release:
 383 ; GCN:        %bb.0
 384 ; GCN-NOT:    ATOMIC_FENCE
 385 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 386 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 387 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 388 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 389 ; GCN:        s_endpgm
 390 ; GFX10:         .amdhsa_kernel system_release
 391 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 392 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 393 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 394 define amdgpu_kernel void @system_release() {
 395 entry:
 396   fence release
 397   ret void
 398 }
 399
 400 ; FUNC-LABEL: {{^}}system_acq_rel:
 401 ; GCN:        %bb.0
 402 ; GCN-NOT:    ATOMIC_FENCE
 403 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 404 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 405 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 406 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 407 ; GFX6:       buffer_wbinvl1{{$}}
 408 ; GFX8:       buffer_wbinvl1_vol{{$}}
 409 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 410 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 411 ; GCN:        s_endpgm
 412 ; GFX10:         .amdhsa_kernel system_acq_rel
 413 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 414 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 415 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 416 define amdgpu_kernel void @system_acq_rel() {
 417 entry:
 418   fence acq_rel
 419   ret void
 420 }
 421
 422 ; FUNC-LABEL: {{^}}system_seq_cst:
 423 ; GCN:        %bb.0
 424 ; GCN-NOT:    ATOMIC_FENCE
 425 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 426 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 427 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 428 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 429 ; GFX6:       buffer_wbinvl1{{$}}
 430 ; GFX8:       buffer_wbinvl1_vol{{$}}
 431 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 432 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 433 ; GCN:        s_endpgm
 434 ; GFX10:         .amdhsa_kernel system_seq_cst
 435 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 436 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 437 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 438 define amdgpu_kernel void @system_seq_cst() {
 439 entry:
 440   fence seq_cst
 441   ret void
 442 }
 443
 444 ; FUNC-LABEL: {{^}}singlethread_acquire:
 445 ; GCN:        %bb.0
 446 ; GCN-NOT:    ATOMIC_FENCE
 447 ; GCN:        s_endpgm
 448 ; GFX10:         .amdhsa_kernel singlethread_acquire
 449 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 450 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 451 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 452 define amdgpu_kernel void @singlethread_acquire() {
 453 entry:
 454   fence syncscope("singlethread") acquire
 455   ret void
 456 }
 457
 458 ; FUNC-LABEL: {{^}}singlethread_release:
 459 ; GCN:        %bb.0
 460 ; GCN-NOT:    ATOMIC_FENCE
 461 ; GCN:        s_endpgm
 462 ; GFX10:         .amdhsa_kernel singlethread_release
 463 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 464 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 465 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 466 define amdgpu_kernel void @singlethread_release() {
 467 entry:
 468   fence syncscope("singlethread") release
 469   ret void
 470 }
 471
 472 ; FUNC-LABEL: {{^}}singlethread_acq_rel:
 473 ; GCN:        %bb.0
 474 ; GCN-NOT:    ATOMIC_FENCE
 475 ; GCN:        s_endpgm
 476 ; GFX10:         .amdhsa_kernel singlethread_acq_rel
 477 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 478 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 479 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 480 define amdgpu_kernel void @singlethread_acq_rel() {
 481 entry:
 482   fence syncscope("singlethread") acq_rel
 483   ret void
 484 }
 485
 486 ; FUNC-LABEL: {{^}}singlethread_seq_cst:
 487 ; GCN:        %bb.0
 488 ; GCN-NOT:    ATOMIC_FENCE
 489 ; GCN:        s_endpgm
 490 ; GFX10:         .amdhsa_kernel singlethread_seq_cst
 491 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 492 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 493 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 494 define amdgpu_kernel void @singlethread_seq_cst() {
 495 entry:
 496   fence syncscope("singlethread") seq_cst
 497   ret void
 498 }
 499
 500 ; FUNC-LABEL: {{^}}agent_acquire:
 501 ; GCN:        %bb.0
 502 ; GCN-NOT:    ATOMIC_FENCE
 503 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 504 ; GFX6-NEXT:  buffer_wbinvl1{{$}}
 505 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 506 ; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
 507 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 508 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 509 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 510 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 511 ; GCN:        s_endpgm
 512 ; GFX10:         .amdhsa_kernel agent_acquire
 513 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 514 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 515 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 516 define amdgpu_kernel void @agent_acquire() {
 517 entry:
 518   fence syncscope("agent") acquire
 519   ret void
 520 }
 521
 522 ; FUNC-LABEL: {{^}}agent_release:
 523 ; GCN:        %bb.0
 524 ; GCN-NOT:    ATOMIC_FENCE
 525 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 526 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 527 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 528 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 529 ; GCN:        s_endpgm
 530 ; GFX10:         .amdhsa_kernel agent_release
 531 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 532 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 533 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 534 define amdgpu_kernel void @agent_release() {
 535 entry:
 536   fence syncscope("agent") release
 537   ret void
 538 }
 539
 540 ; FUNC-LABEL: {{^}}agent_acq_rel:
 541 ; GCN:        %bb.0
 542 ; GCN-NOT:    ATOMIC_FENCE
 543 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 544 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 545 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 546 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 547 ; GFX6:       buffer_wbinvl1{{$}}
 548 ; GFX8:       buffer_wbinvl1_vol{{$}}
 549 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 550 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 551 ; GCN:        s_endpgm
 552 ; GFX10:         .amdhsa_kernel agent_acq_rel
 553 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 554 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 555 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 556 define amdgpu_kernel void @agent_acq_rel() {
 557 entry:
 558   fence syncscope("agent") acq_rel
 559   ret void
 560 }
 561
 562 ; FUNC-LABEL: {{^}}agent_seq_cst:
 563 ; GCN:        %bb.0
 564 ; GCN-NOT:    ATOMIC_FENCE
 565 ; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 566 ; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 567 ; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 568 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 569 ; GFX6:       buffer_wbinvl1{{$}}
 570 ; GFX8:       buffer_wbinvl1_vol{{$}}
 571 ; GFX10-NEXT: buffer_gl0_inv{{$}}
 572 ; GFX10-NEXT: buffer_gl1_inv{{$}}
 573 ; GCN:        s_endpgm
 574 ; GFX10:         .amdhsa_kernel agent_seq_cst
 575 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 576 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 577 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 578 define amdgpu_kernel void @agent_seq_cst() {
 579 entry:
 580   fence syncscope("agent") seq_cst
 581   ret void
 582 }
 583
 584 ; FUNC-LABEL:    {{^}}workgroup_acquire:
 585 ; GCN:           %bb.0
 586 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 587 ; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 588 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 589 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 590 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 591 ; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
 592 ; GFX10CU-NOT:   buffer_gl0_inv{{$}}
 593 ; GCN-NOT:       ATOMIC_FENCE
 594 ; GCN:           s_endpgm
 595 ; GFX10:         .amdhsa_kernel workgroup_acquire
 596 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 597 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 598 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 599 define amdgpu_kernel void @workgroup_acquire() {
 600 entry:
 601   fence syncscope("workgroup") acquire
 602   ret void
 603 }
 604
 605 ; FUNC-LABEL:    {{^}}workgroup_release:
 606 ; GCN:           %bb.0
 607 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 608 ; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 609 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 610 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 611 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 612 ; GFX10-NOT:     buffer_gl0_inv
 613 ; GCN-NOT:       ATOMIC_FENCE
 614 ; GCN:           s_endpgm
 615 ; GFX10:         .amdhsa_kernel workgroup_release
 616 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 617 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 618 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 619 define amdgpu_kernel void @workgroup_release() {
 620 entry:
 621   fence syncscope("workgroup") release
 622   ret void
 623 }
 624
 625 ; FUNC-LABEL:    {{^}}workgroup_acq_rel:
 626 ; GCN:           %bb.0
 627 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 628 ; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 629 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 630 ; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
 631 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 632 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 633 ; GFX10CU-NOT:   buffer_gl0_inv{{$}}
 634 ; GCN-NOT:       ATOMIC_FENCE
 635 ; GCN:           s_endpgm
 636 ; GFX10:         .amdhsa_kernel workgroup_acq_rel
 637 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 638 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 639 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 640 define amdgpu_kernel void @workgroup_acq_rel() {
 641 entry:
 642   fence syncscope("workgroup") acq_rel
 643   ret void
 644 }
 645
 646 ; FUNC-LABEL:    {{^}}workgroup_seq_cst:
 647 ; GCN:           %bb.0
 648 ; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
 649 ; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
 650 ; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
 651 ; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
 652 ; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
 653 ; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
 654 ; GFX10CU-NOT:   buffer_gl0_inv{{$}}
 655 ; GCN-NOT:       ATOMIC_FENCE
 656 ; GCN:           s_endpgm
 657 ; GFX10:         .amdhsa_kernel workgroup_seq_cst
 658 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 659 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 660 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 661 define amdgpu_kernel void @workgroup_seq_cst() {
 662 entry:
 663   fence syncscope("workgroup") seq_cst
 664   ret void
 665 }
 666
 667 ; FUNC-LABEL: {{^}}wavefront_acquire:
 668 ; GCN:        %bb.0
 669 ; GCN-NOT:    ATOMIC_FENCE
 670 ; GCN:        s_endpgm
 671 ; GFX10:         .amdhsa_kernel wavefront_acquire
 672 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 673 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 674 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 675 define amdgpu_kernel void @wavefront_acquire() {
 676 entry:
 677   fence syncscope("wavefront") acquire
 678   ret void
 679 }
 680
 681 ; FUNC-LABEL: {{^}}wavefront_release:
 682 ; GCN:        %bb.0
 683 ; GCN-NOT:    ATOMIC_FENCE
 684 ; GCN:        s_endpgm
 685 ; GFX10:         .amdhsa_kernel wavefront_release
 686 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 687 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 688 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 689 define amdgpu_kernel void @wavefront_release() {
 690 entry:
 691   fence syncscope("wavefront") release
 692   ret void
 693 }
 694
 695 ; FUNC-LABEL: {{^}}wavefront_acq_rel:
 696 ; GCN:        %bb.0
 697 ; GCN-NOT:    ATOMIC_FENCE
 698 ; GCN:        s_endpgm
 699 ; GFX10:         .amdhsa_kernel wavefront_acq_rel
 700 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 701 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 702 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 703 define amdgpu_kernel void @wavefront_acq_rel() {
 704 entry:
 705   fence syncscope("wavefront") acq_rel
 706   ret void
 707 }
 708
 709 ; FUNC-LABEL: {{^}}wavefront_seq_cst:
 710 ; GCN:        %bb.0
 711 ; GCN-NOT:    ATOMIC_FENCE
 712 ; GCN:        s_endpgm
 713 ; GFX10:         .amdhsa_kernel wavefront_seq_cst
 714 ; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
 715 ; GFX10CU:       .amdhsa_workgroup_processor_mode 0
 716 ; GFX10-NOT:     .amdhsa_memory_ordered 0
 717 define amdgpu_kernel void @wavefront_seq_cst() {
 718 entry:
 719   fence syncscope("wavefront") seq_cst
 720   ret void
 721 }