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