llvm/test/CodeGen/AMDGPU/fold-readlane.mir

   1 # RUN: llc -march=amdgcn -run-pass si-fold-operands -verify-machineinstrs %s -o - | FileCheck -check-prefix=GCN %s
   2
   3 # GCN-LABEL: name: fold-imm-readfirstlane{{$}}
   4 # GCN: %1:sreg_32_xm0 = S_MOV_B32 123
   5 ---
   6 name: fold-imm-readfirstlane
   7 tracksRegLiveness: true
   8 body:             |
   9   bb.0:
  10     %0:vgpr_32 = V_MOV_B32_e32 123, implicit $exec
  11     %1:sreg_32_xm0 = V_READFIRSTLANE_B32 %0, implicit $exec
  12     S_NOP 0, implicit %1
  13 ...
  14
  15 # GCN-LABEL: name: fold-imm-readfirstlane-dbgvalue{{$}}
  16 # GCN: %1:sreg_32_xm0 = S_MOV_B32 123
  17 # GCN: DBG_VALUE %0:vgpr_32, 0, 0
  18 ---
  19 name: fold-imm-readfirstlane-dbgvalue
  20 tracksRegLiveness: true
  21 body:             |
  22   bb.0:
  23     %0:vgpr_32 = V_MOV_B32_e32 123, implicit $exec
  24     %1:sreg_32_xm0 = V_READFIRSTLANE_B32 %0, implicit $exec
  25     DBG_VALUE %0, 0, 0
  26     S_NOP 0, implicit %1
  27 ...
  28
  29 # GCN-LABEL: name: fold-imm-readfirstlane-readfirstlane{{$}}
  30 # GCN: %1:sreg_32_xm0 = S_MOV_B32 123
  31 # GCN: %3:sreg_32_xm0 = COPY %1
  32
  33 ---
  34 name: fold-imm-readfirstlane-readfirstlane
  35 tracksRegLiveness: true
  36 body:             |
  37   bb.0:
  38     %0:vgpr_32 = V_MOV_B32_e32 123, implicit $exec
  39     %1:sreg_32_xm0 = V_READFIRSTLANE_B32 %0, implicit $exec
  40     %2:vgpr_32 = COPY %1
  41     %3:sreg_32_xm0 = V_READFIRSTLANE_B32 %2, implicit $exec
  42     S_NOP 0, implicit %3
  43 ...
  44
  45
  46 # GCN-LABEL: name: fold-copy-readfirstlane{{$}}
  47 # GCN: %0:sreg_32_xm0 = COPY $sgpr10
  48 # GCN: %2:sreg_32_xm0 = COPY %0
  49 ---
  50 name: fold-copy-readfirstlane
  51 tracksRegLiveness: true
  52 body:             |
  53   bb.0:
  54     liveins: $sgpr10
  55     %0:sreg_32_xm0 = COPY $sgpr10
  56     %1:vgpr_32 = COPY %0
  57     %2:sreg_32_xm0 = V_READFIRSTLANE_B32 %1, implicit $exec
  58     S_NOP 0, implicit %2
  59 ...
  60
  61 # GCN-LABEL: name: no-fold-copy-readfirstlane-physreg0{{$}}
  62 # GCN: %0:vgpr_32 = COPY $sgpr10
  63 # GCN-NEXT: %1:sreg_32_xm0 = V_READFIRSTLANE_B32 %0, implicit $exec
  64 ---
  65 name: no-fold-copy-readfirstlane-physreg0
  66 tracksRegLiveness: true
  67 body:             |
  68   bb.0:
  69     liveins: $sgpr10
  70     %0:vgpr_32 = COPY $sgpr10
  71     %1:sreg_32_xm0 = V_READFIRSTLANE_B32 %0, implicit $exec
  72
  73 ...
  74
  75 # GCN-LABEL: name: no-fold-copy-readfirstlane-physreg1{{$}}
  76 # GCN: $vgpr0 = COPY $sgpr10
  77 # GCN-NEXT: %0:sreg_32_xm0 = V_READFIRSTLANE_B32 $vgpr0, implicit $exec
  78 ---
  79 name: no-fold-copy-readfirstlane-physreg1
  80 tracksRegLiveness: true
  81 body:             |
  82   bb.0:
  83     liveins: $sgpr10
  84     $vgpr0 = COPY $sgpr10
  85     %0:sreg_32_xm0 = V_READFIRSTLANE_B32 $vgpr0, implicit $exec
  86
  87 ...
  88
  89 # GCN-LABEL: name: no-fold-imm-readfirstlane-physreg{{$}}
  90 # GCN: $vgpr0 = V_MOV_B32_e32 123, implicit $exec
  91 # GCN-NEXT: V_READFIRSTLANE_B32 $vgpr0, implicit $exec
  92
  93 ---
  94 name: no-fold-imm-readfirstlane-physreg
  95 tracksRegLiveness: true
  96 body:             |
  97   bb.0:
  98     $vgpr0 = V_MOV_B32_e32 123, implicit $exec
  99     %0:sreg_32_xm0 = V_READFIRSTLANE_B32 $vgpr0, implicit $exec
 100 ...
 101
 102 # TODO: This could be folded, if the search for exec modifications was
 103 # smarter.
 104
 105 # GCN-LABEL: name: fold-imm-readfirstlane-cross-block{{$}}
 106 # GCN: V_MOV_B32
 107 # GCN: V_READFIRSTLANE_B32
 108 ---
 109 name: fold-imm-readfirstlane-cross-block
 110 tracksRegLiveness: true
 111 body:             |
 112   bb.0:
 113     %0:vgpr_32 = V_MOV_B32_e32 123, implicit $exec
 114
 115   bb.1:
 116     %1:sreg_32_xm0 = V_READFIRSTLANE_B32 %0, implicit $exec
 117 ...
 118
 119 # TODO: This could be folded, if the search for exec modifications was
 120 # smarter.
 121
 122 # GCN-LABEL: name: fold-copy-readfirstlane-cross-block{{$}}
 123 # GCN: V_MOV_B32
 124 # GCN: V_READFIRSTLANE_B32
 125 ---
 126 name: fold-copy-readfirstlane-cross-block
 127 tracksRegLiveness: true
 128 body:             |
 129   bb.0:
 130     liveins: $sgpr12
 131     %0:sreg_32_xm0 = COPY $sgpr12
 132     %1:vgpr_32 = V_MOV_B32_e32 %0, implicit $exec
 133
 134   bb.1:
 135     %2:sreg_32_xm0 = V_READFIRSTLANE_B32 %1, implicit $exec
 136 ...
 137
 138 # GCN-LABEL: name: fold-copy-readfirstlane-cross-block-exec-def{{$}}
 139 # GCN: V_MOV_B32
 140 # GCN: $exec = S_MOV_B64_term
 141 # GCN: V_READFIRSTLANE_B32
 142 ---
 143 name: fold-copy-readfirstlane-cross-block-exec-def
 144 tracksRegLiveness: true
 145 body:             |
 146   bb.0:
 147     liveins: $sgpr10_sgpr11, $sgpr12
 148     %0:sreg_32_xm0 = COPY $sgpr12
 149     %1:vgpr_32 = V_MOV_B32_e32 %0, implicit $exec
 150     $exec = S_MOV_B64_term $sgpr10_sgpr11
 151
 152   bb.1:
 153     %2:sreg_32_xm0 = V_READFIRSTLANE_B32 %1, implicit $exec
 154 ...
 155
 156 # GCN-LABEL: name: fold-copy-readfirstlane-same-block-exec-def{{$}}
 157 # GCN: COPY
 158 # GCN-NEXT: %1:vgpr_32 = COPY %0
 159 # GCN-NEXT: $exec = S_MOV_B64
 160 # GCN-NEXT: V_READFIRSTLANE_B32
 161 ---
 162 name: fold-copy-readfirstlane-same-block-exec-def
 163 tracksRegLiveness: true
 164 body:             |
 165   bb.0:
 166     liveins: $sgpr10_sgpr11, $sgpr12
 167     %0:sreg_32_xm0 = COPY $sgpr12
 168     %1:vgpr_32 = COPY %0, implicit $exec
 169     $exec = S_MOV_B64 $sgpr10_sgpr11
 170     %2:sreg_32_xm0 = V_READFIRSTLANE_B32 %1, implicit $exec
 171
 172 ...
 173
 174 # GCN-LABEL: name: fold-imm-readfirstlane-cross-block-exec-def{{$}}
 175 # GCN: V_MOV_B32
 176 # GCN: $exec = S_MOV_B64
 177 # GCN: V_READFIRSTLANE_B32
 178
 179 ---
 180 name: fold-imm-readfirstlane-cross-block-exec-def
 181 tracksRegLiveness: true
 182 body:             |
 183   bb.0:
 184     liveins: $sgpr10_sgpr11, $sgpr12_sgpr13
 185     %0:vgpr_32 = V_MOV_B32_e32 123, implicit $exec
 186     $exec = S_MOV_B64_term $sgpr10_sgpr11
 187
 188   bb.1:
 189     %1:sreg_32_xm0 = V_READFIRSTLANE_B32 %0, implicit $exec
 190 ...
 191
 192 # GCN-LABEL: name: fold-imm-readfirstlane-same-block-exec-def{{$}}
 193 # GCN: V_MOV_B32
 194 # GCN-NEXT: $exec = S_MOV_B64
 195 # GCN-NEXT: V_READFIRSTLANE_B32
 196 ---
 197 name: fold-imm-readfirstlane-same-block-exec-def
 198 tracksRegLiveness: true
 199 body:             |
 200   bb.0:
 201     liveins: $sgpr10_sgpr11
 202     %0:vgpr_32 = V_MOV_B32_e32 123, implicit $exec
 203     $exec = S_MOV_B64 $sgpr10_sgpr11
 204     %1:sreg_32_xm0 = V_READFIRSTLANE_B32 %0, implicit $exec
 205
 206 ...
 207
 208 # GCN-LABEL: name: fold-sgpr-copy-readfirstlane-same-block-exec-def{{$}}
 209 # GCN: COPY
 210 # GCN-NEXT: $exec = S_MOV_B64
 211 # GCN-NEXT: V_READFIRSTLANE_B32
 212 ---
 213 name: fold-sgpr-copy-readfirstlane-same-block-exec-def
 214 tracksRegLiveness: true
 215 body:             |
 216   bb.0:
 217     liveins: $sgpr10_sgpr11, $sgpr12
 218     %0:vgpr_32 = COPY $sgpr12
 219     $exec = S_MOV_B64 $sgpr10_sgpr11
 220     %1:sreg_32_xm0 = V_READFIRSTLANE_B32 %0, implicit $exec
 221 ...
 222
 223 # GCN-LABEL: name: fold-imm-readfirstlane-user{{$}}
 224 # GCN: %3:sreg_32_xm0 = S_MOV_B32 123
 225 ---
 226 name: fold-imm-readfirstlane-user
 227 tracksRegLiveness: true
 228 body:             |
 229   bb.0:
 230     liveins: $vgpr0, $sgpr0_sgpr1
 231     %0:vgpr_32 = V_MOV_B32_e32 123, implicit $exec
 232     %1:sreg_32_xm0 = V_READFIRSTLANE_B32 %0, implicit $exec
 233     %2:sreg_32_xm0 = COPY %1
 234     %3:sreg_32_xm0 = COPY %2
 235     S_ENDPGM 0, implicit %3
 236 ...
 237
 238 # GCN-LABEL: name: fold-imm-readlane{{$}}
 239 # GCN: %1:sreg_32_xm0 = S_MOV_B32 123
 240 ---
 241 name: fold-imm-readlane
 242 tracksRegLiveness: true
 243 body:             |
 244   bb.0:
 245     liveins: $vgpr0, $sgpr0_sgpr1
 246     %0:vgpr_32 = V_MOV_B32_e32 123, implicit $exec
 247     %1:sreg_32_xm0 = V_READLANE_B32 %0, 0, implicit $exec
 248     S_NOP 0, implicit %1
 249 ...
 250
 251 # GCN-LABEL: name: fold-imm-readlane-src1{{$}}
 252 # GCN: %0:vgpr_32 = COPY $vgpr0
 253 # GCN: V_READLANE_B32 %0, 12, implicit $exec
 254 ---
 255 name: fold-imm-readlane-src1
 256 tracksRegLiveness: true
 257 body:             |
 258   bb.0:
 259     liveins: $vgpr0
 260     %0:vgpr_32 = COPY $vgpr0
 261     %1:sreg_32_xm0 = S_MOV_B32 12
 262     %2:sreg_32_xm0 = V_READLANE_B32 %0, %1, implicit $exec
 263 ...
 264
 265 # Constant for subreg0
 266 # GCN-LABEL: name: fold-imm-readfirstlane-regsequence0{{$}}
 267
 268 # GCN: %0:vgpr_32 = COPY $vgpr0
 269 # GCN-NEXT: %1:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
 270 # GCN-NEXT: %2:vreg_64 = REG_SEQUENCE %0, %subreg.sub0, killed %1, %subreg.sub1
 271 # GCN-NEXT: %3:sgpr_32 = V_READFIRSTLANE_B32 %2.sub0, implicit $exec
 272 # GCN-NEXT: %4:sgpr_32 = S_MOV_B32 0
 273 ---
 274 name: fold-imm-readfirstlane-regsequence0
 275 tracksRegLiveness: true
 276 body:             |
 277   bb.0:
 278     liveins: $vgpr0
 279     %0:vgpr_32 = COPY $vgpr0
 280     %1:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
 281     %2:vreg_64 = REG_SEQUENCE %0:vgpr_32, %subreg.sub0, killed %1:vgpr_32, %subreg.sub1
 282     %3:sgpr_32 = V_READFIRSTLANE_B32 %2.sub0:vreg_64, implicit $exec
 283     %4:sgpr_32 = V_READFIRSTLANE_B32 %2.sub1:vreg_64, implicit $exec
 284     S_NOP 0, implicit %3, implicit %4
 285 ...
 286
 287 # Constant for subreg1
 288 # GCN-LABEL: name: fold-imm-readfirstlane-regsequence1{{$}}
 289 # GCN: %0:vgpr_32 = COPY $vgpr0
 290 # GCN-NEXT: %1:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
 291 # GCN-NEXT: %2:vreg_64 = REG_SEQUENCE %1, %subreg.sub0, killed %0, %subreg.sub1
 292 # GCN-NEXT: %3:sgpr_32 = S_MOV_B32 0
 293 # GCN-NEXT: %4:sgpr_32 = V_READFIRSTLANE_B32 %2.sub1, implicit $exec
 294
 295 ---
 296 name: fold-imm-readfirstlane-regsequence1
 297 tracksRegLiveness: true
 298 body:             |
 299   bb.0:
 300     liveins: $vgpr0
 301     %0:vgpr_32 = COPY $vgpr0
 302     %1:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
 303     %2:vreg_64 = REG_SEQUENCE %1:vgpr_32, %subreg.sub0, killed %0:vgpr_32, %subreg.sub1
 304     %3:sgpr_32 = V_READFIRSTLANE_B32 %2.sub0:vreg_64, implicit $exec
 305     %4:sgpr_32 = V_READFIRSTLANE_B32 %2.sub1:vreg_64, implicit $exec
 306     S_NOP 0, implicit %3, implicit %4
 307 ...
 308
 309 # Different constant regs for each subreg
 310 # GCN-LABEL: name: fold-imm-readfirstlane-regsequence2{{$}}
 311 # GCN: %0:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
 312 # GCN-NEXT: %1:vgpr_32 = V_MOV_B32_e32 1, implicit $exec
 313 # GCN-NEXT: %2:vreg_64 = REG_SEQUENCE %0, %subreg.sub0, killed %1, %subreg.sub1
 314 # GCN-NEXT: %3:sgpr_32 = S_MOV_B32 0
 315 # GCN-NEXT: %4:sgpr_32 = S_MOV_B32 1
 316 ---
 317 name: fold-imm-readfirstlane-regsequence2
 318 tracksRegLiveness: true
 319 body:             |
 320   bb.0:
 321     %0:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
 322     %1:vgpr_32 = V_MOV_B32_e32 1, implicit $exec
 323     %2:vreg_64 = REG_SEQUENCE %0:vgpr_32, %subreg.sub0, killed %1:vgpr_32, %subreg.sub1
 324     %3:sgpr_32 = V_READFIRSTLANE_B32 %2.sub0:vreg_64, implicit $exec
 325     %4:sgpr_32 = V_READFIRSTLANE_B32 %2.sub1:vreg_64, implicit $exec
 326     S_NOP 0, implicit %3, implicit %4
 327 ...
 328
 329 # Same constant reg for each subreg, so there are multiple constant uses
 330 # GCN-LABEL: name: fold-imm-readfirstlane-regsequence3{{$}}
 331 # GCN: %0:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
 332 # GCN-NEXT: %1:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
 333 # GCN-NEXT: %2:vreg_64 = REG_SEQUENCE %0, %subreg.sub0, killed %1, %subreg.sub1
 334 # GCN-NEXT: %3:sgpr_32 = S_MOV_B32 0
 335 # GCN-NEXT: %4:sgpr_32 = S_MOV_B32 0
 336 ---
 337 name: fold-imm-readfirstlane-regsequence3
 338 tracksRegLiveness: true
 339 body:             |
 340   bb.0:
 341     %0:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
 342     %1:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
 343     %2:vreg_64 = REG_SEQUENCE %0:vgpr_32, %subreg.sub0, killed %1:vgpr_32, %subreg.sub1
 344     %3:sgpr_32 = V_READFIRSTLANE_B32 %2.sub0:vreg_64, implicit $exec
 345     %4:sgpr_32 = V_READFIRSTLANE_B32 %2.sub1:vreg_64, implicit $exec
 346     S_NOP 0, implicit %3, implicit %4
 347 ...
 348
 349 # FIXME: This should fold
 350 # GCN-LABEL: name: fold-copy-readfirstlane-regsequence0{{$}}
 351 # GCN: %0:vgpr_32 = COPY $sgpr10
 352 # GCN-NEXT: %1:vgpr_32 = COPY $sgpr11
 353 # GCN-NEXT: %2:vreg_64 = REG_SEQUENCE %0, %subreg.sub0, killed %1, %subreg.sub1
 354 # GCN-NEXT: %3:sgpr_32 = V_READFIRSTLANE_B32 %2.sub0, implicit $exec
 355 # GCN-NEXT: %4:sgpr_32 = V_READFIRSTLANE_B32 %2.sub1, implicit $exec
 356 ---
 357 name: fold-copy-readfirstlane-regsequence0
 358 tracksRegLiveness: true
 359 body:             |
 360   bb.0:
 361     liveins: $sgpr10, $sgpr11
 362     %0:vgpr_32 = COPY $sgpr10
 363     %1:vgpr_32 = COPY $sgpr11
 364     %2:vreg_64 = REG_SEQUENCE %0:vgpr_32, %subreg.sub0, killed %1:vgpr_32, %subreg.sub1
 365     %3:sgpr_32 = V_READFIRSTLANE_B32 %2.sub0:vreg_64, implicit $exec
 366     %4:sgpr_32 = V_READFIRSTLANE_B32 %2.sub1:vreg_64, implicit $exec
 367 ...
 368
 369 # GCN-LABEL: name: fold-copy-readfirstlane-regsequence1{{$}}
 370 # GCN: %0:sreg_32_xm0 = COPY $sgpr10
 371 # GCN-NEXT: %1:sreg_32_xm0 = COPY $sgpr11
 372 # GCN-NEXT: %2:vgpr_32 = COPY %0
 373 # GCN-NEXT: %3:vgpr_32 = COPY %1
 374 # GCN-NEXT: %4:vreg_64 = REG_SEQUENCE %2, %subreg.sub0, killed %3, %subreg.sub1
 375 # GCN-NEXT: %5:sgpr_32 = V_READFIRSTLANE_B32 %4.sub0, implicit $exec
 376 # GCN-NEXT: %6:sgpr_32 = V_READFIRSTLANE_B32 %4.sub1, implicit $exec
 377 ---
 378 name: fold-copy-readfirstlane-regsequence1
 379 tracksRegLiveness: true
 380 body:             |
 381   bb.0:
 382     liveins: $sgpr10, $sgpr11
 383     %0:sreg_32_xm0 = COPY $sgpr10
 384     %1:sreg_32_xm0 = COPY $sgpr11
 385     %2:vgpr_32 = COPY %0
 386     %3:vgpr_32 = COPY %1
 387     %4:vreg_64 = REG_SEQUENCE %2:vgpr_32, %subreg.sub0, killed %3:vgpr_32, %subreg.sub1
 388     %5:sgpr_32 = V_READFIRSTLANE_B32 %4.sub0:vreg_64, implicit $exec
 389     %6:sgpr_32 = V_READFIRSTLANE_B32 %4.sub1:vreg_64, implicit $exec
 390 ...