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

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