test/tools/llvm-mca/X86/Barcelona/load-throughput.s

   1 # NOTE: Assertions have been autogenerated by utils/update_mca_test_checks.py
   2 # RUN: llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -scheduler-stats -dispatch-stats -iterations=100 -timeline -timeline-max-iterations=1 < %s | FileCheck %s
   3
   4 # LLVM-MCA-BEGIN
   5 movb (%rax), %spl
   6 movb (%rcx), %bpl
   7 movb (%rdx), %sil
   8 movb (%rbx), %dil
   9 # LLVM-MCA-END
  10
  11 # LLVM-MCA-BEGIN
  12 movw (%rax), %sp
  13 movw (%rcx), %bp
  14 movw (%rdx), %si
  15 movw (%rbx), %di
  16 # LLVM-MCA-END
  17
  18 # LLVM-MCA-BEGIN
  19 movl (%rax), %esp
  20 movl (%rcx), %ebp
  21 movl (%rdx), %esi
  22 movl (%rbx), %edi
  23 # LLVM-MCA-END
  24
  25 # LLVM-MCA-BEGIN
  26 movq (%rax), %rsp
  27 movq (%rcx), %rbp
  28 movq (%rdx), %rsi
  29 movq (%rbx), %rdi
  30 # LLVM-MCA-END
  31
  32 # LLVM-MCA-BEGIN
  33 movd (%rax), %mm0
  34 movd (%rcx), %mm1
  35 movd (%rdx), %mm2
  36 movd (%rbx), %mm3
  37 # LLVM-MCA-END
  38
  39 # LLVM-MCA-BEGIN
  40 movaps (%rax), %xmm0
  41 movaps (%rcx), %xmm1
  42 movaps (%rdx), %xmm2
  43 movaps (%rbx), %xmm3
  44 # LLVM-MCA-END
  45
  46 # CHECK:      [0] Code Region
  47
  48 # CHECK:      Iterations:        100
  49 # CHECK-NEXT: Instructions:      400
  50 # CHECK-NEXT: Total Cycles:      207
  51 # CHECK-NEXT: Total uOps:        400
  52
  53 # CHECK:      Dispatch Width:    4
  54 # CHECK-NEXT: uOps Per Cycle:    1.93
  55 # CHECK-NEXT: IPC:               1.93
  56 # CHECK-NEXT: Block RThroughput: 2.0
  57
  58 # CHECK:      Instruction Info:
  59 # CHECK-NEXT: [1]: #uOps
  60 # CHECK-NEXT: [2]: Latency
  61 # CHECK-NEXT: [3]: RThroughput
  62 # CHECK-NEXT: [4]: MayLoad
  63 # CHECK-NEXT: [5]: MayStore
  64 # CHECK-NEXT: [6]: HasSideEffects (U)
  65
  66 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
  67 # CHECK-NEXT:  1      5     0.50    *                   movb    (%rax), %spl
  68 # CHECK-NEXT:  1      5     0.50    *                   movb    (%rcx), %bpl
  69 # CHECK-NEXT:  1      5     0.50    *                   movb    (%rdx), %sil
  70 # CHECK-NEXT:  1      5     0.50    *                   movb    (%rbx), %dil
  71
  72 # CHECK:      Dynamic Dispatch Stall Cycles:
  73 # CHECK-NEXT: RAT     - Register unavailable:                      0
  74 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
  75 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            147  (71.0%)
  76 # CHECK-NEXT: LQ      - Load queue full:                           0
  77 # CHECK-NEXT: SQ      - Store queue full:                          0
  78 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
  79
  80 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
  81 # CHECK-NEXT: [# dispatched], [# cycles]
  82 # CHECK-NEXT:  0,              33  (15.9%)
  83 # CHECK-NEXT:  2,              148  (71.5%)
  84 # CHECK-NEXT:  4,              26  (12.6%)
  85
  86 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
  87 # CHECK-NEXT: [# issued], [# cycles]
  88 # CHECK-NEXT:  0,          7  (3.4%)
  89 # CHECK-NEXT:  2,          200  (96.6%)
  90
  91 # CHECK:      Scheduler's queue usage:
  92 # CHECK-NEXT: [1] Resource name.
  93 # CHECK-NEXT: [2] Average number of used buffer entries.
  94 # CHECK-NEXT: [3] Maximum number of used buffer entries.
  95 # CHECK-NEXT: [4] Total number of buffer entries.
  96
  97 # CHECK:       [1]            [2]        [3]        [4]
  98 # CHECK-NEXT: SBPortAny        45         54         54
  99
 100 # CHECK:      Resources:
 101 # CHECK-NEXT: [0]   - SBDivider
 102 # CHECK-NEXT: [1]   - SBFPDivider
 103 # CHECK-NEXT: [2]   - SBPort0
 104 # CHECK-NEXT: [3]   - SBPort1
 105 # CHECK-NEXT: [4]   - SBPort4
 106 # CHECK-NEXT: [5]   - SBPort5
 107 # CHECK-NEXT: [6.0] - SBPort23
 108 # CHECK-NEXT: [6.1] - SBPort23
 109
 110 # CHECK:      Resource pressure per iteration:
 111 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 112 # CHECK-NEXT:  -      -      -      -      -      -     2.00   2.00
 113
 114 # CHECK:      Resource pressure by instruction:
 115 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 116 # CHECK-NEXT:  -      -      -      -      -      -      -     1.00   movb      (%rax), %spl
 117 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movb      (%rcx), %bpl
 118 # CHECK-NEXT:  -      -      -      -      -      -      -     1.00   movb      (%rdx), %sil
 119 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movb      (%rbx), %dil
 120
 121 # CHECK:      Timeline view:
 122 # CHECK-NEXT: Index     012345678
 123
 124 # CHECK:      [0,0]     DeeeeeER.   movb        (%rax), %spl
 125 # CHECK-NEXT: [0,1]     DeeeeeER.   movb        (%rcx), %bpl
 126 # CHECK-NEXT: [0,2]     D=eeeeeER   movb        (%rdx), %sil
 127 # CHECK-NEXT: [0,3]     D=eeeeeER   movb        (%rbx), %dil
 128
 129 # CHECK:      Average Wait times (based on the timeline view):
 130 # CHECK-NEXT: [0]: Executions
 131 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 132 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 133 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 134
 135 # CHECK:            [0]    [1]    [2]    [3]
 136 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movb (%rax), %spl
 137 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movb (%rcx), %bpl
 138 # CHECK-NEXT: 2.     1     2.0    2.0    0.0       movb (%rdx), %sil
 139 # CHECK-NEXT: 3.     1     2.0    2.0    0.0       movb (%rbx), %dil
 140
 141 # CHECK:      [1] Code Region
 142
 143 # CHECK:      Iterations:        100
 144 # CHECK-NEXT: Instructions:      400
 145 # CHECK-NEXT: Total Cycles:      207
 146 # CHECK-NEXT: Total uOps:        400
 147
 148 # CHECK:      Dispatch Width:    4
 149 # CHECK-NEXT: uOps Per Cycle:    1.93
 150 # CHECK-NEXT: IPC:               1.93
 151 # CHECK-NEXT: Block RThroughput: 2.0
 152
 153 # CHECK:      Instruction Info:
 154 # CHECK-NEXT: [1]: #uOps
 155 # CHECK-NEXT: [2]: Latency
 156 # CHECK-NEXT: [3]: RThroughput
 157 # CHECK-NEXT: [4]: MayLoad
 158 # CHECK-NEXT: [5]: MayStore
 159 # CHECK-NEXT: [6]: HasSideEffects (U)
 160
 161 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 162 # CHECK-NEXT:  1      5     0.50    *                   movw    (%rax), %sp
 163 # CHECK-NEXT:  1      5     0.50    *                   movw    (%rcx), %bp
 164 # CHECK-NEXT:  1      5     0.50    *                   movw    (%rdx), %si
 165 # CHECK-NEXT:  1      5     0.50    *                   movw    (%rbx), %di
 166
 167 # CHECK:      Dynamic Dispatch Stall Cycles:
 168 # CHECK-NEXT: RAT     - Register unavailable:                      0
 169 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 170 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            147  (71.0%)
 171 # CHECK-NEXT: LQ      - Load queue full:                           0
 172 # CHECK-NEXT: SQ      - Store queue full:                          0
 173 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 174
 175 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 176 # CHECK-NEXT: [# dispatched], [# cycles]
 177 # CHECK-NEXT:  0,              33  (15.9%)
 178 # CHECK-NEXT:  2,              148  (71.5%)
 179 # CHECK-NEXT:  4,              26  (12.6%)
 180
 181 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 182 # CHECK-NEXT: [# issued], [# cycles]
 183 # CHECK-NEXT:  0,          7  (3.4%)
 184 # CHECK-NEXT:  2,          200  (96.6%)
 185
 186 # CHECK:      Scheduler's queue usage:
 187 # CHECK-NEXT: [1] Resource name.
 188 # CHECK-NEXT: [2] Average number of used buffer entries.
 189 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 190 # CHECK-NEXT: [4] Total number of buffer entries.
 191
 192 # CHECK:       [1]            [2]        [3]        [4]
 193 # CHECK-NEXT: SBPortAny        45         54         54
 194
 195 # CHECK:      Resources:
 196 # CHECK-NEXT: [0]   - SBDivider
 197 # CHECK-NEXT: [1]   - SBFPDivider
 198 # CHECK-NEXT: [2]   - SBPort0
 199 # CHECK-NEXT: [3]   - SBPort1
 200 # CHECK-NEXT: [4]   - SBPort4
 201 # CHECK-NEXT: [5]   - SBPort5
 202 # CHECK-NEXT: [6.0] - SBPort23
 203 # CHECK-NEXT: [6.1] - SBPort23
 204
 205 # CHECK:      Resource pressure per iteration:
 206 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 207 # CHECK-NEXT:  -      -      -      -      -      -     2.00   2.00
 208
 209 # CHECK:      Resource pressure by instruction:
 210 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 211 # CHECK-NEXT:  -      -      -      -      -      -      -     1.00   movw      (%rax), %sp
 212 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movw      (%rcx), %bp
 213 # CHECK-NEXT:  -      -      -      -      -      -      -     1.00   movw      (%rdx), %si
 214 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movw      (%rbx), %di
 215
 216 # CHECK:      Timeline view:
 217 # CHECK-NEXT: Index     012345678
 218
 219 # CHECK:      [0,0]     DeeeeeER.   movw        (%rax), %sp
 220 # CHECK-NEXT: [0,1]     DeeeeeER.   movw        (%rcx), %bp
 221 # CHECK-NEXT: [0,2]     D=eeeeeER   movw        (%rdx), %si
 222 # CHECK-NEXT: [0,3]     D=eeeeeER   movw        (%rbx), %di
 223
 224 # CHECK:      Average Wait times (based on the timeline view):
 225 # CHECK-NEXT: [0]: Executions
 226 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 227 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 228 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 229
 230 # CHECK:            [0]    [1]    [2]    [3]
 231 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movw (%rax), %sp
 232 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movw (%rcx), %bp
 233 # CHECK-NEXT: 2.     1     2.0    2.0    0.0       movw (%rdx), %si
 234 # CHECK-NEXT: 3.     1     2.0    2.0    0.0       movw (%rbx), %di
 235
 236 # CHECK:      [2] Code Region
 237
 238 # CHECK:      Iterations:        100
 239 # CHECK-NEXT: Instructions:      400
 240 # CHECK-NEXT: Total Cycles:      207
 241 # CHECK-NEXT: Total uOps:        400
 242
 243 # CHECK:      Dispatch Width:    4
 244 # CHECK-NEXT: uOps Per Cycle:    1.93
 245 # CHECK-NEXT: IPC:               1.93
 246 # CHECK-NEXT: Block RThroughput: 2.0
 247
 248 # CHECK:      Instruction Info:
 249 # CHECK-NEXT: [1]: #uOps
 250 # CHECK-NEXT: [2]: Latency
 251 # CHECK-NEXT: [3]: RThroughput
 252 # CHECK-NEXT: [4]: MayLoad
 253 # CHECK-NEXT: [5]: MayStore
 254 # CHECK-NEXT: [6]: HasSideEffects (U)
 255
 256 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 257 # CHECK-NEXT:  1      5     0.50    *                   movl    (%rax), %esp
 258 # CHECK-NEXT:  1      5     0.50    *                   movl    (%rcx), %ebp
 259 # CHECK-NEXT:  1      5     0.50    *                   movl    (%rdx), %esi
 260 # CHECK-NEXT:  1      5     0.50    *                   movl    (%rbx), %edi
 261
 262 # CHECK:      Dynamic Dispatch Stall Cycles:
 263 # CHECK-NEXT: RAT     - Register unavailable:                      0
 264 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 265 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            147  (71.0%)
 266 # CHECK-NEXT: LQ      - Load queue full:                           0
 267 # CHECK-NEXT: SQ      - Store queue full:                          0
 268 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 269
 270 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 271 # CHECK-NEXT: [# dispatched], [# cycles]
 272 # CHECK-NEXT:  0,              33  (15.9%)
 273 # CHECK-NEXT:  2,              148  (71.5%)
 274 # CHECK-NEXT:  4,              26  (12.6%)
 275
 276 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 277 # CHECK-NEXT: [# issued], [# cycles]
 278 # CHECK-NEXT:  0,          7  (3.4%)
 279 # CHECK-NEXT:  2,          200  (96.6%)
 280
 281 # CHECK:      Scheduler's queue usage:
 282 # CHECK-NEXT: [1] Resource name.
 283 # CHECK-NEXT: [2] Average number of used buffer entries.
 284 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 285 # CHECK-NEXT: [4] Total number of buffer entries.
 286
 287 # CHECK:       [1]            [2]        [3]        [4]
 288 # CHECK-NEXT: SBPortAny        45         54         54
 289
 290 # CHECK:      Resources:
 291 # CHECK-NEXT: [0]   - SBDivider
 292 # CHECK-NEXT: [1]   - SBFPDivider
 293 # CHECK-NEXT: [2]   - SBPort0
 294 # CHECK-NEXT: [3]   - SBPort1
 295 # CHECK-NEXT: [4]   - SBPort4
 296 # CHECK-NEXT: [5]   - SBPort5
 297 # CHECK-NEXT: [6.0] - SBPort23
 298 # CHECK-NEXT: [6.1] - SBPort23
 299
 300 # CHECK:      Resource pressure per iteration:
 301 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 302 # CHECK-NEXT:  -      -      -      -      -      -     2.00   2.00
 303
 304 # CHECK:      Resource pressure by instruction:
 305 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 306 # CHECK-NEXT:  -      -      -      -      -      -      -     1.00   movl      (%rax), %esp
 307 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movl      (%rcx), %ebp
 308 # CHECK-NEXT:  -      -      -      -      -      -      -     1.00   movl      (%rdx), %esi
 309 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movl      (%rbx), %edi
 310
 311 # CHECK:      Timeline view:
 312 # CHECK-NEXT: Index     012345678
 313
 314 # CHECK:      [0,0]     DeeeeeER.   movl        (%rax), %esp
 315 # CHECK-NEXT: [0,1]     DeeeeeER.   movl        (%rcx), %ebp
 316 # CHECK-NEXT: [0,2]     D=eeeeeER   movl        (%rdx), %esi
 317 # CHECK-NEXT: [0,3]     D=eeeeeER   movl        (%rbx), %edi
 318
 319 # CHECK:      Average Wait times (based on the timeline view):
 320 # CHECK-NEXT: [0]: Executions
 321 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 322 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 323 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 324
 325 # CHECK:            [0]    [1]    [2]    [3]
 326 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movl (%rax), %esp
 327 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movl (%rcx), %ebp
 328 # CHECK-NEXT: 2.     1     2.0    2.0    0.0       movl (%rdx), %esi
 329 # CHECK-NEXT: 3.     1     2.0    2.0    0.0       movl (%rbx), %edi
 330
 331 # CHECK:      [3] Code Region
 332
 333 # CHECK:      Iterations:        100
 334 # CHECK-NEXT: Instructions:      400
 335 # CHECK-NEXT: Total Cycles:      207
 336 # CHECK-NEXT: Total uOps:        400
 337
 338 # CHECK:      Dispatch Width:    4
 339 # CHECK-NEXT: uOps Per Cycle:    1.93
 340 # CHECK-NEXT: IPC:               1.93
 341 # CHECK-NEXT: Block RThroughput: 2.0
 342
 343 # CHECK:      Instruction Info:
 344 # CHECK-NEXT: [1]: #uOps
 345 # CHECK-NEXT: [2]: Latency
 346 # CHECK-NEXT: [3]: RThroughput
 347 # CHECK-NEXT: [4]: MayLoad
 348 # CHECK-NEXT: [5]: MayStore
 349 # CHECK-NEXT: [6]: HasSideEffects (U)
 350
 351 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 352 # CHECK-NEXT:  1      5     0.50    *                   movq    (%rax), %rsp
 353 # CHECK-NEXT:  1      5     0.50    *                   movq    (%rcx), %rbp
 354 # CHECK-NEXT:  1      5     0.50    *                   movq    (%rdx), %rsi
 355 # CHECK-NEXT:  1      5     0.50    *                   movq    (%rbx), %rdi
 356
 357 # CHECK:      Dynamic Dispatch Stall Cycles:
 358 # CHECK-NEXT: RAT     - Register unavailable:                      0
 359 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 360 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            147  (71.0%)
 361 # CHECK-NEXT: LQ      - Load queue full:                           0
 362 # CHECK-NEXT: SQ      - Store queue full:                          0
 363 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 364
 365 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 366 # CHECK-NEXT: [# dispatched], [# cycles]
 367 # CHECK-NEXT:  0,              33  (15.9%)
 368 # CHECK-NEXT:  2,              148  (71.5%)
 369 # CHECK-NEXT:  4,              26  (12.6%)
 370
 371 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 372 # CHECK-NEXT: [# issued], [# cycles]
 373 # CHECK-NEXT:  0,          7  (3.4%)
 374 # CHECK-NEXT:  2,          200  (96.6%)
 375
 376 # CHECK:      Scheduler's queue usage:
 377 # CHECK-NEXT: [1] Resource name.
 378 # CHECK-NEXT: [2] Average number of used buffer entries.
 379 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 380 # CHECK-NEXT: [4] Total number of buffer entries.
 381
 382 # CHECK:       [1]            [2]        [3]        [4]
 383 # CHECK-NEXT: SBPortAny        45         54         54
 384
 385 # CHECK:      Resources:
 386 # CHECK-NEXT: [0]   - SBDivider
 387 # CHECK-NEXT: [1]   - SBFPDivider
 388 # CHECK-NEXT: [2]   - SBPort0
 389 # CHECK-NEXT: [3]   - SBPort1
 390 # CHECK-NEXT: [4]   - SBPort4
 391 # CHECK-NEXT: [5]   - SBPort5
 392 # CHECK-NEXT: [6.0] - SBPort23
 393 # CHECK-NEXT: [6.1] - SBPort23
 394
 395 # CHECK:      Resource pressure per iteration:
 396 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 397 # CHECK-NEXT:  -      -      -      -      -      -     2.00   2.00
 398
 399 # CHECK:      Resource pressure by instruction:
 400 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 401 # CHECK-NEXT:  -      -      -      -      -      -      -     1.00   movq      (%rax), %rsp
 402 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movq      (%rcx), %rbp
 403 # CHECK-NEXT:  -      -      -      -      -      -      -     1.00   movq      (%rdx), %rsi
 404 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movq      (%rbx), %rdi
 405
 406 # CHECK:      Timeline view:
 407 # CHECK-NEXT: Index     012345678
 408
 409 # CHECK:      [0,0]     DeeeeeER.   movq        (%rax), %rsp
 410 # CHECK-NEXT: [0,1]     DeeeeeER.   movq        (%rcx), %rbp
 411 # CHECK-NEXT: [0,2]     D=eeeeeER   movq        (%rdx), %rsi
 412 # CHECK-NEXT: [0,3]     D=eeeeeER   movq        (%rbx), %rdi
 413
 414 # CHECK:      Average Wait times (based on the timeline view):
 415 # CHECK-NEXT: [0]: Executions
 416 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 417 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 418 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 419
 420 # CHECK:            [0]    [1]    [2]    [3]
 421 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movq (%rax), %rsp
 422 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movq (%rcx), %rbp
 423 # CHECK-NEXT: 2.     1     2.0    2.0    0.0       movq (%rdx), %rsi
 424 # CHECK-NEXT: 3.     1     2.0    2.0    0.0       movq (%rbx), %rdi
 425
 426 # CHECK:      [4] Code Region
 427
 428 # CHECK:      Iterations:        100
 429 # CHECK-NEXT: Instructions:      400
 430 # CHECK-NEXT: Total Cycles:      207
 431 # CHECK-NEXT: Total uOps:        400
 432
 433 # CHECK:      Dispatch Width:    4
 434 # CHECK-NEXT: uOps Per Cycle:    1.93
 435 # CHECK-NEXT: IPC:               1.93
 436 # CHECK-NEXT: Block RThroughput: 2.0
 437
 438 # CHECK:      Instruction Info:
 439 # CHECK-NEXT: [1]: #uOps
 440 # CHECK-NEXT: [2]: Latency
 441 # CHECK-NEXT: [3]: RThroughput
 442 # CHECK-NEXT: [4]: MayLoad
 443 # CHECK-NEXT: [5]: MayStore
 444 # CHECK-NEXT: [6]: HasSideEffects (U)
 445
 446 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 447 # CHECK-NEXT:  1      5     0.50    *                   movd    (%rax), %mm0
 448 # CHECK-NEXT:  1      5     0.50    *                   movd    (%rcx), %mm1
 449 # CHECK-NEXT:  1      5     0.50    *                   movd    (%rdx), %mm2
 450 # CHECK-NEXT:  1      5     0.50    *                   movd    (%rbx), %mm3
 451
 452 # CHECK:      Dynamic Dispatch Stall Cycles:
 453 # CHECK-NEXT: RAT     - Register unavailable:                      0
 454 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 455 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            147  (71.0%)
 456 # CHECK-NEXT: LQ      - Load queue full:                           0
 457 # CHECK-NEXT: SQ      - Store queue full:                          0
 458 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 459
 460 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 461 # CHECK-NEXT: [# dispatched], [# cycles]
 462 # CHECK-NEXT:  0,              33  (15.9%)
 463 # CHECK-NEXT:  2,              148  (71.5%)
 464 # CHECK-NEXT:  4,              26  (12.6%)
 465
 466 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 467 # CHECK-NEXT: [# issued], [# cycles]
 468 # CHECK-NEXT:  0,          7  (3.4%)
 469 # CHECK-NEXT:  2,          200  (96.6%)
 470
 471 # CHECK:      Scheduler's queue usage:
 472 # CHECK-NEXT: [1] Resource name.
 473 # CHECK-NEXT: [2] Average number of used buffer entries.
 474 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 475 # CHECK-NEXT: [4] Total number of buffer entries.
 476
 477 # CHECK:       [1]            [2]        [3]        [4]
 478 # CHECK-NEXT: SBPortAny        45         54         54
 479
 480 # CHECK:      Resources:
 481 # CHECK-NEXT: [0]   - SBDivider
 482 # CHECK-NEXT: [1]   - SBFPDivider
 483 # CHECK-NEXT: [2]   - SBPort0
 484 # CHECK-NEXT: [3]   - SBPort1
 485 # CHECK-NEXT: [4]   - SBPort4
 486 # CHECK-NEXT: [5]   - SBPort5
 487 # CHECK-NEXT: [6.0] - SBPort23
 488 # CHECK-NEXT: [6.1] - SBPort23
 489
 490 # CHECK:      Resource pressure per iteration:
 491 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 492 # CHECK-NEXT:  -      -      -      -      -      -     2.00   2.00
 493
 494 # CHECK:      Resource pressure by instruction:
 495 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 496 # CHECK-NEXT:  -      -      -      -      -      -      -     1.00   movd      (%rax), %mm0
 497 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movd      (%rcx), %mm1
 498 # CHECK-NEXT:  -      -      -      -      -      -      -     1.00   movd      (%rdx), %mm2
 499 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movd      (%rbx), %mm3
 500
 501 # CHECK:      Timeline view:
 502 # CHECK-NEXT: Index     012345678
 503
 504 # CHECK:      [0,0]     DeeeeeER.   movd        (%rax), %mm0
 505 # CHECK-NEXT: [0,1]     DeeeeeER.   movd        (%rcx), %mm1
 506 # CHECK-NEXT: [0,2]     D=eeeeeER   movd        (%rdx), %mm2
 507 # CHECK-NEXT: [0,3]     D=eeeeeER   movd        (%rbx), %mm3
 508
 509 # CHECK:      Average Wait times (based on the timeline view):
 510 # CHECK-NEXT: [0]: Executions
 511 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 512 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 513 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 514
 515 # CHECK:            [0]    [1]    [2]    [3]
 516 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movd (%rax), %mm0
 517 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movd (%rcx), %mm1
 518 # CHECK-NEXT: 2.     1     2.0    2.0    0.0       movd (%rdx), %mm2
 519 # CHECK-NEXT: 3.     1     2.0    2.0    0.0       movd (%rbx), %mm3
 520
 521 # CHECK:      [5] Code Region
 522
 523 # CHECK:      Iterations:        100
 524 # CHECK-NEXT: Instructions:      400
 525 # CHECK-NEXT: Total Cycles:      208
 526 # CHECK-NEXT: Total uOps:        400
 527
 528 # CHECK:      Dispatch Width:    4
 529 # CHECK-NEXT: uOps Per Cycle:    1.92
 530 # CHECK-NEXT: IPC:               1.92
 531 # CHECK-NEXT: Block RThroughput: 2.0
 532
 533 # CHECK:      Instruction Info:
 534 # CHECK-NEXT: [1]: #uOps
 535 # CHECK-NEXT: [2]: Latency
 536 # CHECK-NEXT: [3]: RThroughput
 537 # CHECK-NEXT: [4]: MayLoad
 538 # CHECK-NEXT: [5]: MayStore
 539 # CHECK-NEXT: [6]: HasSideEffects (U)
 540
 541 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 542 # CHECK-NEXT:  1      6     0.50    *                   movaps  (%rax), %xmm0
 543 # CHECK-NEXT:  1      6     0.50    *                   movaps  (%rcx), %xmm1
 544 # CHECK-NEXT:  1      6     0.50    *                   movaps  (%rdx), %xmm2
 545 # CHECK-NEXT:  1      6     0.50    *                   movaps  (%rbx), %xmm3
 546
 547 # CHECK:      Dynamic Dispatch Stall Cycles:
 548 # CHECK-NEXT: RAT     - Register unavailable:                      0
 549 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 550 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            147  (70.7%)
 551 # CHECK-NEXT: LQ      - Load queue full:                           0
 552 # CHECK-NEXT: SQ      - Store queue full:                          0
 553 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 554
 555 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 556 # CHECK-NEXT: [# dispatched], [# cycles]
 557 # CHECK-NEXT:  0,              34  (16.3%)
 558 # CHECK-NEXT:  2,              148  (71.2%)
 559 # CHECK-NEXT:  4,              26  (12.5%)
 560
 561 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 562 # CHECK-NEXT: [# issued], [# cycles]
 563 # CHECK-NEXT:  0,          8  (3.8%)
 564 # CHECK-NEXT:  2,          200  (96.2%)
 565
 566 # CHECK:      Scheduler's queue usage:
 567 # CHECK-NEXT: [1] Resource name.
 568 # CHECK-NEXT: [2] Average number of used buffer entries.
 569 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 570 # CHECK-NEXT: [4] Total number of buffer entries.
 571
 572 # CHECK:       [1]            [2]        [3]        [4]
 573 # CHECK-NEXT: SBPortAny        45         54         54
 574
 575 # CHECK:      Resources:
 576 # CHECK-NEXT: [0]   - SBDivider
 577 # CHECK-NEXT: [1]   - SBFPDivider
 578 # CHECK-NEXT: [2]   - SBPort0
 579 # CHECK-NEXT: [3]   - SBPort1
 580 # CHECK-NEXT: [4]   - SBPort4
 581 # CHECK-NEXT: [5]   - SBPort5
 582 # CHECK-NEXT: [6.0] - SBPort23
 583 # CHECK-NEXT: [6.1] - SBPort23
 584
 585 # CHECK:      Resource pressure per iteration:
 586 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 587 # CHECK-NEXT:  -      -      -      -      -      -     2.00   2.00
 588
 589 # CHECK:      Resource pressure by instruction:
 590 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 591 # CHECK-NEXT:  -      -      -      -      -      -      -     1.00   movaps    (%rax), %xmm0
 592 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movaps    (%rcx), %xmm1
 593 # CHECK-NEXT:  -      -      -      -      -      -      -     1.00   movaps    (%rdx), %xmm2
 594 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movaps    (%rbx), %xmm3
 595
 596 # CHECK:      Timeline view:
 597 # CHECK-NEXT: Index     0123456789
 598
 599 # CHECK:      [0,0]     DeeeeeeER.   movaps     (%rax), %xmm0
 600 # CHECK-NEXT: [0,1]     DeeeeeeER.   movaps     (%rcx), %xmm1
 601 # CHECK-NEXT: [0,2]     D=eeeeeeER   movaps     (%rdx), %xmm2
 602 # CHECK-NEXT: [0,3]     D=eeeeeeER   movaps     (%rbx), %xmm3
 603
 604 # CHECK:      Average Wait times (based on the timeline view):
 605 # CHECK-NEXT: [0]: Executions
 606 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 607 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 608 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 609
 610 # CHECK:            [0]    [1]    [2]    [3]
 611 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movaps       (%rax), %xmm0
 612 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movaps       (%rcx), %xmm1
 613 # CHECK-NEXT: 2.     1     2.0    2.0    0.0       movaps       (%rdx), %xmm2
 614 # CHECK-NEXT: 3.     1     2.0    2.0    0.0       movaps       (%rbx), %xmm3