llvm/test/CodeGen/AMDGPU/vmem-to-salu-hazard.mir

   1 # RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs -run-pass post-RA-hazard-rec -o - %s | FileCheck -check-prefixes=GCN,GFX10 %s
   2 # RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs -run-pass post-RA-hazard-rec -o - %s | FileCheck -check-prefixes=GCN %s
   3
   4 # GCN-LABEL: name: vmem_write_sgpr
   5 # GCN:        BUFFER_LOAD_DWORD_OFFEN
   6 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
   7 # GCN-NEXT:   S_MOV_B32
   8 ---
   9 name:            vmem_write_sgpr
  10 body:             |
  11   bb.0:
  12     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
  13     $sgpr4 = IMPLICIT_DEF
  14     $vgpr0 = IMPLICIT_DEF
  15     $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
  16     $sgpr0 = S_MOV_B32 0
  17 ...
  18 # GCN-LABEL: name: vmem_write_exec
  19 # GCN:        BUFFER_STORE_DWORD_OFFEN_exact
  20 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
  21 # GCN-NEXT:   S_MOV_B32
  22 ---
  23 name:            vmem_write_exec
  24 body:             |
  25   bb.0:
  26     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
  27     $sgpr4 = IMPLICIT_DEF
  28     $vgpr0 = IMPLICIT_DEF
  29     $vgpr1 = IMPLICIT_DEF
  30     BUFFER_STORE_DWORD_OFFEN_exact killed renamable $vgpr0, renamable $vgpr1, renamable $sgpr0_sgpr1_sgpr2_sgpr3, 0, 0, 0, 0, implicit $exec
  31     $exec_lo = S_MOV_B32 -1
  32 ...
  33 # GCN-LABEL: name: vmem_write_sgpr_chain
  34 # GCN:        BUFFER_LOAD_DWORD_OFFEN
  35 # GCN-NEXT:   S_MOV_B32
  36 # GCN-NEXT:   S_MOV_B32
  37 # GCN-NEXT:   S_MOV_B32
  38 # GCN-NEXT:   S_MOV_B32
  39 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
  40 # GCN-NEXT:   S_MOV_B32
  41 ---
  42 name:            vmem_write_sgpr_chain
  43 body:             |
  44   bb.0:
  45     successors:
  46     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
  47     $sgpr4 = IMPLICIT_DEF
  48     $vgpr0 = IMPLICIT_DEF
  49     $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
  50     $sgpr5 = S_MOV_B32 $sgpr0
  51     $sgpr6 = S_MOV_B32 $sgpr1
  52     $sgpr7 = S_MOV_B32 $sgpr2
  53     $sgpr8 = S_MOV_B32 $sgpr3
  54     $sgpr0 = S_MOV_B32 0
  55 ...
  56 # GCN-LABEL: name: vmem_smem_write_sgpr
  57 # GCN:        BUFFER_LOAD_DWORD_OFFEN
  58 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
  59 # GCN-NEXT:   S_LOAD_DWORD_IMM
  60 ---
  61 name:            vmem_smem_write_sgpr
  62 body:             |
  63   bb.0:
  64     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
  65     $sgpr4 = IMPLICIT_DEF
  66     $vgpr0 = IMPLICIT_DEF
  67     $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
  68     $sgpr0 = S_LOAD_DWORD_IMM $sgpr0_sgpr1, 0, 0
  69 ...
  70 # GCN-LABEL: name: vmem_snop_write_sgpr
  71 # GCN:        BUFFER_LOAD_DWORD_OFFEN
  72 # GCN-NEXT:   S_NOP
  73 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
  74 # GCN-NEXT:   S_MOV_B32
  75 ---
  76 name:            vmem_snop_write_sgpr
  77 body:             |
  78   bb.0:
  79     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
  80     $sgpr4 = IMPLICIT_DEF
  81     $vgpr0 = IMPLICIT_DEF
  82     $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
  83     S_NOP 0
  84     $sgpr0 = S_MOV_B32 0
  85 ...
  86 # GCN-LABEL: name: vmem_valu_write_sgpr
  87 # GCN:      BUFFER_LOAD_DWORD_OFFEN
  88 # GCN-NEXT: V_ADD_F32
  89 # GCN-NEXT: S_MOV_B32
  90 ---
  91 name:            vmem_valu_write_sgpr
  92 body:             |
  93   bb.0:
  94     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
  95     $sgpr4 = IMPLICIT_DEF
  96     $vgpr0 = IMPLICIT_DEF
  97     $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
  98     $vgpr2 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
  99     $sgpr0 = S_MOV_B32 0
 100 ...
 101 # GCN-LABEL: name: vmem_swait0_write_sgpr
 102 # GCN:      BUFFER_LOAD_DWORD_OFFEN
 103 # GCN-NEXT: S_WAITCNT
 104 # GCN-NEXT: S_MOV_B32
 105 ---
 106 name:            vmem_swait0_write_sgpr
 107 body:             |
 108   bb.0:
 109     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
 110     $sgpr4 = IMPLICIT_DEF
 111     $vgpr0 = IMPLICIT_DEF
 112     $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
 113     S_WAITCNT 0
 114     $sgpr0 = S_MOV_B32 0
 115 ...
 116 # GCN-LABEL: name: vmem_swait_any_write_sgpr
 117 # GCN:        BUFFER_LOAD_DWORD_OFFEN
 118 # GCN-NEXT:   S_WAITCNT
 119 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
 120 # GCN-NEXT:   S_MOV_B32
 121 ---
 122 name:            vmem_swait_any_write_sgpr
 123 body:             |
 124   bb.0:
 125     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
 126     $sgpr4 = IMPLICIT_DEF
 127     $vgpr0 = IMPLICIT_DEF
 128     $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
 129     S_WAITCNT 1
 130     $sgpr0 = S_MOV_B32 0
 131 ...
 132 # GCN-LABEL: name: vmem_write_exec_impread
 133 # GCN:        BUFFER_LOAD_DWORD_OFFEN
 134 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
 135 # GCN-NEXT:   S_MOV_B64
 136 ---
 137 name:            vmem_write_exec_impread
 138 body:             |
 139   bb.0:
 140     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
 141     $sgpr4 = IMPLICIT_DEF
 142     $vgpr0 = IMPLICIT_DEF
 143     $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
 144     $exec = S_MOV_B64 7
 145 ...
 146 # GCN-LABEL: name: vmem_write_exec_expread
 147 # GCN:        BUFFER_LOAD_DWORD_OFFEN
 148 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
 149 # GCN-NEXT:   S_MOV_B64
 150 ---
 151 name:            vmem_write_exec_expread
 152 body:             |
 153   bb.0:
 154     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
 155     $vgpr0 = IMPLICIT_DEF
 156     $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $exec_lo, 0, 0, 0, implicit $exec
 157     $exec = S_MOV_B64 7
 158 ...
 159 # GCN-LABEL: name: ds_write_m0
 160 # GCN:        DS_READ_B32
 161 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
 162 # GCN-NEXT:   S_MOV_B32
 163 ---
 164 name:            ds_write_m0
 165 body:             |
 166   bb.0:
 167     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
 168     $sgpr4 = IMPLICIT_DEF
 169     $vgpr0 = IMPLICIT_DEF
 170     $vgpr1 = DS_READ_B32 $vgpr0, 0, 0, implicit $m0, implicit $exec
 171     $m0 = S_MOV_B32 7
 172 ...
 173 # GCN-LABEL: name: vmem_write_sgpr_fall_through
 174 # GCN:        BUFFER_LOAD_DWORD_OFFEN
 175 # GCN:        bb.1:
 176 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
 177 # GCN-NEXT:   S_MOV_B32
 178 ---
 179 name:            vmem_write_sgpr_fall_through
 180 body:             |
 181   bb.0:
 182     successors: %bb.1
 183     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
 184     $sgpr4 = IMPLICIT_DEF
 185     $vgpr0 = IMPLICIT_DEF
 186     $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
 187
 188   bb.1:
 189     $sgpr0 = S_MOV_B32 0
 190 ...
 191 # GCN-LABEL: name: vmem_write_sgpr_branch
 192 # GCN:        BUFFER_LOAD_DWORD_OFFEN
 193 # GCN-NEXT:   S_BRANCH
 194 # GCN:        bb.1:
 195 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
 196 # GCN-NEXT:   S_MOV_B32
 197 ---
 198 name:            vmem_write_sgpr_branch
 199 body:             |
 200   bb.0:
 201     successors: %bb.1
 202     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
 203     $sgpr4 = IMPLICIT_DEF
 204     $vgpr0 = IMPLICIT_DEF
 205     $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
 206     S_BRANCH %bb.1
 207
 208   bb.1:
 209     $sgpr0 = S_MOV_B32 0
 210 ...
 211 # GCN-LABEL: name: vmem_write_sgpr_branch_around
 212 # GCN:        BUFFER_LOAD_DWORD_OFFEN
 213 # GCN-NEXT:   S_BRANCH
 214 # GCN:        bb.2:
 215 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
 216 # GCN-NEXT:   S_MOV_B32
 217 ---
 218 name:            vmem_write_sgpr_branch_around
 219 body:             |
 220   bb.0:
 221     successors: %bb.2
 222     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
 223     $sgpr4 = IMPLICIT_DEF
 224     $vgpr0 = IMPLICIT_DEF
 225     $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
 226     S_BRANCH %bb.2
 227
 228   bb.1:
 229     successors: %bb.2
 230     S_WAITCNT 0
 231
 232   bb.2:
 233     $sgpr0 = S_MOV_B32 0
 234 ...
 235 # GCN-LABEL: name: vmem_write_sgpr_cbranch_around
 236 # GCN:        BUFFER_LOAD_DWORD_OFFEN
 237 # GCN-NEXT:   S_CBRANCH
 238 # GCN-NEXT:   S_BRANCH
 239 # GCN:        bb.1:
 240 # GCN:        S_WAITCNT
 241 # GCN:        V_ADD_CO_U32
 242 # GCN:        bb.2:
 243 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
 244 # GCN-NEXT:   S_MOV_B32
 245 ---
 246 name:            vmem_write_sgpr_cbranch_around
 247 body:             |
 248   bb.0:
 249     successors: %bb.1, %bb.2
 250     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
 251     $sgpr4 = IMPLICIT_DEF
 252     $vgpr0 = IMPLICIT_DEF
 253     $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
 254     S_CBRANCH_SCC0 %bb.2, implicit $scc
 255     S_BRANCH %bb.1
 256
 257   bb.1:
 258     successors: %bb.2
 259     S_WAITCNT 0
 260     $vgpr2, $vcc_lo = V_ADD_CO_U32_e64 $vgpr1, $vgpr1, 0, implicit $exec
 261     S_BRANCH %bb.2
 262
 263   bb.2:
 264     $sgpr0 = S_MOV_B32 0
 265 ...
 266 # GCN-LABEL: name: vmem_write_sgpr_branch_backedge
 267 # GCN:        $vgpr0 = IMPLICIT_DEF
 268 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
 269 # GCN-NEXT:   S_MOV_B32
 270 ---
 271 name:            vmem_write_sgpr_branch_backedge
 272 body:             |
 273   bb.0:
 274     successors: %bb.1
 275     $sgpr0_sgpr1_sgpr2_sgpr3 = IMPLICIT_DEF
 276     $sgpr4 = IMPLICIT_DEF
 277     $vgpr0 = IMPLICIT_DEF
 278     $sgpr0 = S_MOV_B32 0
 279
 280   bb.1:
 281     $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
 282     S_BRANCH %bb.0
 283 ...
 284 # GCN-LABEL: name: ds_write_exec
 285 # GCN:        DS_WRITE_B32_gfx9
 286 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
 287 # GCN-NEXT:   S_MOV_B32
 288 ---
 289 name:            ds_write_exec
 290 body:             |
 291   bb.0:
 292     $vgpr0 = IMPLICIT_DEF
 293     $vgpr1 = IMPLICIT_DEF
 294     DS_WRITE_B32_gfx9 $vgpr0, $vgpr1, 0, 0, implicit $exec
 295     $exec_lo = S_MOV_B32 -1
 296 ...
 297 # GCN-LABEL: name: vmem_scratch_exec
 298 # GCN:        SCRATCH_LOAD_DWORD
 299 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
 300 # GCN-NEXT:   S_MOV_B32
 301 ---
 302 name:            vmem_scratch_exec
 303 body:             |
 304   bb.0:
 305     $vgpr0 = IMPLICIT_DEF
 306     $vgpr1 = SCRATCH_LOAD_DWORD undef $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
 307     $exec_lo = S_MOV_B32 -1
 308 ...
 309 # GCN-LABEL: name: vmem_flat_exec
 310 # GCN:        FLAT_LOAD_DWORD
 311 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
 312 # GCN-NEXT:   S_MOV_B32
 313 ---
 314 name:            vmem_flat_exec
 315 body:             |
 316   bb.0:
 317     $vgpr0 = IMPLICIT_DEF
 318     $vgpr1 = IMPLICIT_DEF
 319     $vgpr2 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
 320     $exec_lo = S_MOV_B32 -1
 321 ...
 322 # GCN-LABEL: name: vmem_global_exec
 323 # GCN:        GLOBAL_LOAD_DWORD
 324 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
 325 # GCN-NEXT:   S_MOV_B32
 326 ---
 327 name:            vmem_global_exec
 328 body:             |
 329   bb.0:
 330     $vgpr0 = IMPLICIT_DEF
 331     $vgpr1 = IMPLICIT_DEF
 332     $vgpr2 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec
 333     $exec_lo = S_MOV_B32 -1
 334 ...
 335 # GCN-LABEL: name: vmem_global_atomic_exec
 336 # GCN:        GLOBAL_ATOMIC_ADD_RTN
 337 # GFX10-NEXT: S_WAITCNT_DEPCTR 65507
 338 # GCN-NEXT:   S_MOV_B32
 339 ---
 340 name:            vmem_global_atomic_exec
 341 body:             |
 342   bb.0:
 343     $vgpr0 = IMPLICIT_DEF
 344     $vgpr1 = IMPLICIT_DEF
 345     $vgpr2 = IMPLICIT_DEF
 346     $vgpr3 = GLOBAL_ATOMIC_ADD_RTN $vgpr0_vgpr1, $vgpr2, 0, 1, implicit $exec :: (load store syncscope("agent") seq_cst (s32), addrspace 1)
 347     $exec_lo = S_MOV_B32 -1
 348 ...