test/tools/llvm-mca/X86/Barcelona/store-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  -noalias=true < %s | FileCheck %s
   3
   4 # LLVM-MCA-BEGIN
   5 movb %spl, (%rax)
   6 movb %bpl, (%rcx)
   7 movb %sil, (%rdx)
   8 movb %dil, (%rbx)
   9 # LLVM-MCA-END
  10
  11 # LLVM-MCA-BEGIN
  12 movw %sp, (%rax)
  13 movw %bp, (%rcx)
  14 movw %si, (%rdx)
  15 movw %di, (%rbx)
  16 # LLVM-MCA-END
  17
  18 # LLVM-MCA-BEGIN
  19 movl %esp, (%rax)
  20 movl %ebp, (%rcx)
  21 movl %esi, (%rdx)
  22 movl %edi, (%rbx)
  23 # LLVM-MCA-END
  24
  25 # LLVM-MCA-BEGIN
  26 movq %rsp, (%rax)
  27 movq %rbp, (%rcx)
  28 movq %rsi, (%rdx)
  29 movq %rdi, (%rbx)
  30 # LLVM-MCA-END
  31
  32 # LLVM-MCA-BEGIN
  33 movd %mm0, (%rax)
  34 movd %mm1, (%rcx)
  35 movd %mm2, (%rdx)
  36 movd %mm3, (%rbx)
  37 # LLVM-MCA-END
  38
  39 # LLVM-MCA-BEGIN
  40 movaps %xmm0, (%rax)
  41 movaps %xmm1, (%rcx)
  42 movaps %xmm2, (%rdx)
  43 movaps %xmm3, (%rbx)
  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:      403
  51 # CHECK-NEXT: Total uOps:        400
  52
  53 # CHECK:      Dispatch Width:    4
  54 # CHECK-NEXT: uOps Per Cycle:    0.99
  55 # CHECK-NEXT: IPC:               0.99
  56 # CHECK-NEXT: Block RThroughput: 4.0
  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      1     1.00           *            movb    %spl, (%rax)
  68 # CHECK-NEXT:  1      1     1.00           *            movb    %bpl, (%rcx)
  69 # CHECK-NEXT:  1      1     1.00           *            movb    %sil, (%rdx)
  70 # CHECK-NEXT:  1      1     1.00           *            movb    %dil, (%rbx)
  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:                            329  (81.6%)
  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,              56  (13.9%)
  83 # CHECK-NEXT:  1,              329  (81.6%)
  84 # CHECK-NEXT:  3,              1  (0.2%)
  85 # CHECK-NEXT:  4,              17  (4.2%)
  86
  87 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
  88 # CHECK-NEXT: [# issued], [# cycles]
  89 # CHECK-NEXT:  0,          3  (0.7%)
  90 # CHECK-NEXT:  1,          400  (99.3%)
  91
  92 # CHECK:      Scheduler's queue usage:
  93 # CHECK-NEXT: [1] Resource name.
  94 # CHECK-NEXT: [2] Average number of used buffer entries.
  95 # CHECK-NEXT: [3] Maximum number of used buffer entries.
  96 # CHECK-NEXT: [4] Total number of buffer entries.
  97
  98 # CHECK:       [1]            [2]        [3]        [4]
  99 # CHECK-NEXT: SBPortAny        49         54         54
 100
 101 # CHECK:      Resources:
 102 # CHECK-NEXT: [0]   - SBDivider
 103 # CHECK-NEXT: [1]   - SBFPDivider
 104 # CHECK-NEXT: [2]   - SBPort0
 105 # CHECK-NEXT: [3]   - SBPort1
 106 # CHECK-NEXT: [4]   - SBPort4
 107 # CHECK-NEXT: [5]   - SBPort5
 108 # CHECK-NEXT: [6.0] - SBPort23
 109 # CHECK-NEXT: [6.1] - SBPort23
 110
 111 # CHECK:      Resource pressure per iteration:
 112 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 113 # CHECK-NEXT:  -      -      -      -     4.00    -     2.00   2.00
 114
 115 # CHECK:      Resource pressure by instruction:
 116 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 117 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movb      %spl, (%rax)
 118 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movb      %bpl, (%rcx)
 119 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movb      %sil, (%rdx)
 120 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movb      %dil, (%rbx)
 121
 122 # CHECK:      Timeline view:
 123 # CHECK-NEXT: Index     0123456
 124
 125 # CHECK:      [0,0]     DeER ..   movb  %spl, (%rax)
 126 # CHECK-NEXT: [0,1]     D=eER..   movb  %bpl, (%rcx)
 127 # CHECK-NEXT: [0,2]     D==eER.   movb  %sil, (%rdx)
 128 # CHECK-NEXT: [0,3]     D===eER   movb  %dil, (%rbx)
 129
 130 # CHECK:      Average Wait times (based on the timeline view):
 131 # CHECK-NEXT: [0]: Executions
 132 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 133 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 134 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 135
 136 # CHECK:            [0]    [1]    [2]    [3]
 137 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movb %spl, (%rax)
 138 # CHECK-NEXT: 1.     1     2.0    0.0    0.0       movb %bpl, (%rcx)
 139 # CHECK-NEXT: 2.     1     3.0    0.0    0.0       movb %sil, (%rdx)
 140 # CHECK-NEXT: 3.     1     4.0    0.0    0.0       movb %dil, (%rbx)
 141
 142 # CHECK:      [1] Code Region
 143
 144 # CHECK:      Iterations:        100
 145 # CHECK-NEXT: Instructions:      400
 146 # CHECK-NEXT: Total Cycles:      403
 147 # CHECK-NEXT: Total uOps:        400
 148
 149 # CHECK:      Dispatch Width:    4
 150 # CHECK-NEXT: uOps Per Cycle:    0.99
 151 # CHECK-NEXT: IPC:               0.99
 152 # CHECK-NEXT: Block RThroughput: 4.0
 153
 154 # CHECK:      Instruction Info:
 155 # CHECK-NEXT: [1]: #uOps
 156 # CHECK-NEXT: [2]: Latency
 157 # CHECK-NEXT: [3]: RThroughput
 158 # CHECK-NEXT: [4]: MayLoad
 159 # CHECK-NEXT: [5]: MayStore
 160 # CHECK-NEXT: [6]: HasSideEffects (U)
 161
 162 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 163 # CHECK-NEXT:  1      1     1.00           *            movw    %sp, (%rax)
 164 # CHECK-NEXT:  1      1     1.00           *            movw    %bp, (%rcx)
 165 # CHECK-NEXT:  1      1     1.00           *            movw    %si, (%rdx)
 166 # CHECK-NEXT:  1      1     1.00           *            movw    %di, (%rbx)
 167
 168 # CHECK:      Dynamic Dispatch Stall Cycles:
 169 # CHECK-NEXT: RAT     - Register unavailable:                      0
 170 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 171 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            329  (81.6%)
 172 # CHECK-NEXT: LQ      - Load queue full:                           0
 173 # CHECK-NEXT: SQ      - Store queue full:                          0
 174 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 175
 176 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 177 # CHECK-NEXT: [# dispatched], [# cycles]
 178 # CHECK-NEXT:  0,              56  (13.9%)
 179 # CHECK-NEXT:  1,              329  (81.6%)
 180 # CHECK-NEXT:  3,              1  (0.2%)
 181 # CHECK-NEXT:  4,              17  (4.2%)
 182
 183 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 184 # CHECK-NEXT: [# issued], [# cycles]
 185 # CHECK-NEXT:  0,          3  (0.7%)
 186 # CHECK-NEXT:  1,          400  (99.3%)
 187
 188 # CHECK:      Scheduler's queue usage:
 189 # CHECK-NEXT: [1] Resource name.
 190 # CHECK-NEXT: [2] Average number of used buffer entries.
 191 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 192 # CHECK-NEXT: [4] Total number of buffer entries.
 193
 194 # CHECK:       [1]            [2]        [3]        [4]
 195 # CHECK-NEXT: SBPortAny        49         54         54
 196
 197 # CHECK:      Resources:
 198 # CHECK-NEXT: [0]   - SBDivider
 199 # CHECK-NEXT: [1]   - SBFPDivider
 200 # CHECK-NEXT: [2]   - SBPort0
 201 # CHECK-NEXT: [3]   - SBPort1
 202 # CHECK-NEXT: [4]   - SBPort4
 203 # CHECK-NEXT: [5]   - SBPort5
 204 # CHECK-NEXT: [6.0] - SBPort23
 205 # CHECK-NEXT: [6.1] - SBPort23
 206
 207 # CHECK:      Resource pressure per iteration:
 208 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 209 # CHECK-NEXT:  -      -      -      -     4.00    -     2.00   2.00
 210
 211 # CHECK:      Resource pressure by instruction:
 212 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 213 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movw      %sp, (%rax)
 214 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movw      %bp, (%rcx)
 215 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movw      %si, (%rdx)
 216 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movw      %di, (%rbx)
 217
 218 # CHECK:      Timeline view:
 219 # CHECK-NEXT: Index     0123456
 220
 221 # CHECK:      [0,0]     DeER ..   movw  %sp, (%rax)
 222 # CHECK-NEXT: [0,1]     D=eER..   movw  %bp, (%rcx)
 223 # CHECK-NEXT: [0,2]     D==eER.   movw  %si, (%rdx)
 224 # CHECK-NEXT: [0,3]     D===eER   movw  %di, (%rbx)
 225
 226 # CHECK:      Average Wait times (based on the timeline view):
 227 # CHECK-NEXT: [0]: Executions
 228 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 229 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 230 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 231
 232 # CHECK:            [0]    [1]    [2]    [3]
 233 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movw %sp, (%rax)
 234 # CHECK-NEXT: 1.     1     2.0    0.0    0.0       movw %bp, (%rcx)
 235 # CHECK-NEXT: 2.     1     3.0    0.0    0.0       movw %si, (%rdx)
 236 # CHECK-NEXT: 3.     1     4.0    0.0    0.0       movw %di, (%rbx)
 237
 238 # CHECK:      [2] Code Region
 239
 240 # CHECK:      Iterations:        100
 241 # CHECK-NEXT: Instructions:      400
 242 # CHECK-NEXT: Total Cycles:      403
 243 # CHECK-NEXT: Total uOps:        400
 244
 245 # CHECK:      Dispatch Width:    4
 246 # CHECK-NEXT: uOps Per Cycle:    0.99
 247 # CHECK-NEXT: IPC:               0.99
 248 # CHECK-NEXT: Block RThroughput: 4.0
 249
 250 # CHECK:      Instruction Info:
 251 # CHECK-NEXT: [1]: #uOps
 252 # CHECK-NEXT: [2]: Latency
 253 # CHECK-NEXT: [3]: RThroughput
 254 # CHECK-NEXT: [4]: MayLoad
 255 # CHECK-NEXT: [5]: MayStore
 256 # CHECK-NEXT: [6]: HasSideEffects (U)
 257
 258 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 259 # CHECK-NEXT:  1      1     1.00           *            movl    %esp, (%rax)
 260 # CHECK-NEXT:  1      1     1.00           *            movl    %ebp, (%rcx)
 261 # CHECK-NEXT:  1      1     1.00           *            movl    %esi, (%rdx)
 262 # CHECK-NEXT:  1      1     1.00           *            movl    %edi, (%rbx)
 263
 264 # CHECK:      Dynamic Dispatch Stall Cycles:
 265 # CHECK-NEXT: RAT     - Register unavailable:                      0
 266 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 267 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            329  (81.6%)
 268 # CHECK-NEXT: LQ      - Load queue full:                           0
 269 # CHECK-NEXT: SQ      - Store queue full:                          0
 270 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 271
 272 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 273 # CHECK-NEXT: [# dispatched], [# cycles]
 274 # CHECK-NEXT:  0,              56  (13.9%)
 275 # CHECK-NEXT:  1,              329  (81.6%)
 276 # CHECK-NEXT:  3,              1  (0.2%)
 277 # CHECK-NEXT:  4,              17  (4.2%)
 278
 279 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 280 # CHECK-NEXT: [# issued], [# cycles]
 281 # CHECK-NEXT:  0,          3  (0.7%)
 282 # CHECK-NEXT:  1,          400  (99.3%)
 283
 284 # CHECK:      Scheduler's queue usage:
 285 # CHECK-NEXT: [1] Resource name.
 286 # CHECK-NEXT: [2] Average number of used buffer entries.
 287 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 288 # CHECK-NEXT: [4] Total number of buffer entries.
 289
 290 # CHECK:       [1]            [2]        [3]        [4]
 291 # CHECK-NEXT: SBPortAny        49         54         54
 292
 293 # CHECK:      Resources:
 294 # CHECK-NEXT: [0]   - SBDivider
 295 # CHECK-NEXT: [1]   - SBFPDivider
 296 # CHECK-NEXT: [2]   - SBPort0
 297 # CHECK-NEXT: [3]   - SBPort1
 298 # CHECK-NEXT: [4]   - SBPort4
 299 # CHECK-NEXT: [5]   - SBPort5
 300 # CHECK-NEXT: [6.0] - SBPort23
 301 # CHECK-NEXT: [6.1] - SBPort23
 302
 303 # CHECK:      Resource pressure per iteration:
 304 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 305 # CHECK-NEXT:  -      -      -      -     4.00    -     2.00   2.00
 306
 307 # CHECK:      Resource pressure by instruction:
 308 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 309 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movl      %esp, (%rax)
 310 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movl      %ebp, (%rcx)
 311 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movl      %esi, (%rdx)
 312 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movl      %edi, (%rbx)
 313
 314 # CHECK:      Timeline view:
 315 # CHECK-NEXT: Index     0123456
 316
 317 # CHECK:      [0,0]     DeER ..   movl  %esp, (%rax)
 318 # CHECK-NEXT: [0,1]     D=eER..   movl  %ebp, (%rcx)
 319 # CHECK-NEXT: [0,2]     D==eER.   movl  %esi, (%rdx)
 320 # CHECK-NEXT: [0,3]     D===eER   movl  %edi, (%rbx)
 321
 322 # CHECK:      Average Wait times (based on the timeline view):
 323 # CHECK-NEXT: [0]: Executions
 324 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 325 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 326 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 327
 328 # CHECK:            [0]    [1]    [2]    [3]
 329 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movl %esp, (%rax)
 330 # CHECK-NEXT: 1.     1     2.0    0.0    0.0       movl %ebp, (%rcx)
 331 # CHECK-NEXT: 2.     1     3.0    0.0    0.0       movl %esi, (%rdx)
 332 # CHECK-NEXT: 3.     1     4.0    0.0    0.0       movl %edi, (%rbx)
 333
 334 # CHECK:      [3] Code Region
 335
 336 # CHECK:      Iterations:        100
 337 # CHECK-NEXT: Instructions:      400
 338 # CHECK-NEXT: Total Cycles:      403
 339 # CHECK-NEXT: Total uOps:        400
 340
 341 # CHECK:      Dispatch Width:    4
 342 # CHECK-NEXT: uOps Per Cycle:    0.99
 343 # CHECK-NEXT: IPC:               0.99
 344 # CHECK-NEXT: Block RThroughput: 4.0
 345
 346 # CHECK:      Instruction Info:
 347 # CHECK-NEXT: [1]: #uOps
 348 # CHECK-NEXT: [2]: Latency
 349 # CHECK-NEXT: [3]: RThroughput
 350 # CHECK-NEXT: [4]: MayLoad
 351 # CHECK-NEXT: [5]: MayStore
 352 # CHECK-NEXT: [6]: HasSideEffects (U)
 353
 354 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 355 # CHECK-NEXT:  1      1     1.00           *            movq    %rsp, (%rax)
 356 # CHECK-NEXT:  1      1     1.00           *            movq    %rbp, (%rcx)
 357 # CHECK-NEXT:  1      1     1.00           *            movq    %rsi, (%rdx)
 358 # CHECK-NEXT:  1      1     1.00           *            movq    %rdi, (%rbx)
 359
 360 # CHECK:      Dynamic Dispatch Stall Cycles:
 361 # CHECK-NEXT: RAT     - Register unavailable:                      0
 362 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 363 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            329  (81.6%)
 364 # CHECK-NEXT: LQ      - Load queue full:                           0
 365 # CHECK-NEXT: SQ      - Store queue full:                          0
 366 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 367
 368 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 369 # CHECK-NEXT: [# dispatched], [# cycles]
 370 # CHECK-NEXT:  0,              56  (13.9%)
 371 # CHECK-NEXT:  1,              329  (81.6%)
 372 # CHECK-NEXT:  3,              1  (0.2%)
 373 # CHECK-NEXT:  4,              17  (4.2%)
 374
 375 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 376 # CHECK-NEXT: [# issued], [# cycles]
 377 # CHECK-NEXT:  0,          3  (0.7%)
 378 # CHECK-NEXT:  1,          400  (99.3%)
 379
 380 # CHECK:      Scheduler's queue usage:
 381 # CHECK-NEXT: [1] Resource name.
 382 # CHECK-NEXT: [2] Average number of used buffer entries.
 383 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 384 # CHECK-NEXT: [4] Total number of buffer entries.
 385
 386 # CHECK:       [1]            [2]        [3]        [4]
 387 # CHECK-NEXT: SBPortAny        49         54         54
 388
 389 # CHECK:      Resources:
 390 # CHECK-NEXT: [0]   - SBDivider
 391 # CHECK-NEXT: [1]   - SBFPDivider
 392 # CHECK-NEXT: [2]   - SBPort0
 393 # CHECK-NEXT: [3]   - SBPort1
 394 # CHECK-NEXT: [4]   - SBPort4
 395 # CHECK-NEXT: [5]   - SBPort5
 396 # CHECK-NEXT: [6.0] - SBPort23
 397 # CHECK-NEXT: [6.1] - SBPort23
 398
 399 # CHECK:      Resource pressure per iteration:
 400 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 401 # CHECK-NEXT:  -      -      -      -     4.00    -     2.00   2.00
 402
 403 # CHECK:      Resource pressure by instruction:
 404 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 405 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movq      %rsp, (%rax)
 406 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movq      %rbp, (%rcx)
 407 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movq      %rsi, (%rdx)
 408 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movq      %rdi, (%rbx)
 409
 410 # CHECK:      Timeline view:
 411 # CHECK-NEXT: Index     0123456
 412
 413 # CHECK:      [0,0]     DeER ..   movq  %rsp, (%rax)
 414 # CHECK-NEXT: [0,1]     D=eER..   movq  %rbp, (%rcx)
 415 # CHECK-NEXT: [0,2]     D==eER.   movq  %rsi, (%rdx)
 416 # CHECK-NEXT: [0,3]     D===eER   movq  %rdi, (%rbx)
 417
 418 # CHECK:      Average Wait times (based on the timeline view):
 419 # CHECK-NEXT: [0]: Executions
 420 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 421 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 422 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 423
 424 # CHECK:            [0]    [1]    [2]    [3]
 425 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movq %rsp, (%rax)
 426 # CHECK-NEXT: 1.     1     2.0    0.0    0.0       movq %rbp, (%rcx)
 427 # CHECK-NEXT: 2.     1     3.0    0.0    0.0       movq %rsi, (%rdx)
 428 # CHECK-NEXT: 3.     1     4.0    0.0    0.0       movq %rdi, (%rbx)
 429
 430 # CHECK:      [4] Code Region
 431
 432 # CHECK:      Iterations:        100
 433 # CHECK-NEXT: Instructions:      400
 434 # CHECK-NEXT: Total Cycles:      403
 435 # CHECK-NEXT: Total uOps:        400
 436
 437 # CHECK:      Dispatch Width:    4
 438 # CHECK-NEXT: uOps Per Cycle:    0.99
 439 # CHECK-NEXT: IPC:               0.99
 440 # CHECK-NEXT: Block RThroughput: 4.0
 441
 442 # CHECK:      Instruction Info:
 443 # CHECK-NEXT: [1]: #uOps
 444 # CHECK-NEXT: [2]: Latency
 445 # CHECK-NEXT: [3]: RThroughput
 446 # CHECK-NEXT: [4]: MayLoad
 447 # CHECK-NEXT: [5]: MayStore
 448 # CHECK-NEXT: [6]: HasSideEffects (U)
 449
 450 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 451 # CHECK-NEXT:  1      1     1.00           *      U     movd    %mm0, (%rax)
 452 # CHECK-NEXT:  1      1     1.00           *      U     movd    %mm1, (%rcx)
 453 # CHECK-NEXT:  1      1     1.00           *      U     movd    %mm2, (%rdx)
 454 # CHECK-NEXT:  1      1     1.00           *      U     movd    %mm3, (%rbx)
 455
 456 # CHECK:      Dynamic Dispatch Stall Cycles:
 457 # CHECK-NEXT: RAT     - Register unavailable:                      0
 458 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 459 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            329  (81.6%)
 460 # CHECK-NEXT: LQ      - Load queue full:                           0
 461 # CHECK-NEXT: SQ      - Store queue full:                          0
 462 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 463
 464 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 465 # CHECK-NEXT: [# dispatched], [# cycles]
 466 # CHECK-NEXT:  0,              56  (13.9%)
 467 # CHECK-NEXT:  1,              329  (81.6%)
 468 # CHECK-NEXT:  3,              1  (0.2%)
 469 # CHECK-NEXT:  4,              17  (4.2%)
 470
 471 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 472 # CHECK-NEXT: [# issued], [# cycles]
 473 # CHECK-NEXT:  0,          3  (0.7%)
 474 # CHECK-NEXT:  1,          400  (99.3%)
 475
 476 # CHECK:      Scheduler's queue usage:
 477 # CHECK-NEXT: [1] Resource name.
 478 # CHECK-NEXT: [2] Average number of used buffer entries.
 479 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 480 # CHECK-NEXT: [4] Total number of buffer entries.
 481
 482 # CHECK:       [1]            [2]        [3]        [4]
 483 # CHECK-NEXT: SBPortAny        49         54         54
 484
 485 # CHECK:      Resources:
 486 # CHECK-NEXT: [0]   - SBDivider
 487 # CHECK-NEXT: [1]   - SBFPDivider
 488 # CHECK-NEXT: [2]   - SBPort0
 489 # CHECK-NEXT: [3]   - SBPort1
 490 # CHECK-NEXT: [4]   - SBPort4
 491 # CHECK-NEXT: [5]   - SBPort5
 492 # CHECK-NEXT: [6.0] - SBPort23
 493 # CHECK-NEXT: [6.1] - SBPort23
 494
 495 # CHECK:      Resource pressure per iteration:
 496 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 497 # CHECK-NEXT:  -      -      -      -     4.00    -     2.00   2.00
 498
 499 # CHECK:      Resource pressure by instruction:
 500 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 501 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movd      %mm0, (%rax)
 502 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movd      %mm1, (%rcx)
 503 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movd      %mm2, (%rdx)
 504 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movd      %mm3, (%rbx)
 505
 506 # CHECK:      Timeline view:
 507 # CHECK-NEXT: Index     0123456
 508
 509 # CHECK:      [0,0]     DeER ..   movd  %mm0, (%rax)
 510 # CHECK-NEXT: [0,1]     D=eER..   movd  %mm1, (%rcx)
 511 # CHECK-NEXT: [0,2]     D==eER.   movd  %mm2, (%rdx)
 512 # CHECK-NEXT: [0,3]     D===eER   movd  %mm3, (%rbx)
 513
 514 # CHECK:      Average Wait times (based on the timeline view):
 515 # CHECK-NEXT: [0]: Executions
 516 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 517 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 518 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 519
 520 # CHECK:            [0]    [1]    [2]    [3]
 521 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movd %mm0, (%rax)
 522 # CHECK-NEXT: 1.     1     2.0    0.0    0.0       movd %mm1, (%rcx)
 523 # CHECK-NEXT: 2.     1     3.0    0.0    0.0       movd %mm2, (%rdx)
 524 # CHECK-NEXT: 3.     1     4.0    0.0    0.0       movd %mm3, (%rbx)
 525
 526 # CHECK:      [5] Code Region
 527
 528 # CHECK:      Iterations:        100
 529 # CHECK-NEXT: Instructions:      400
 530 # CHECK-NEXT: Total Cycles:      403
 531 # CHECK-NEXT: Total uOps:        400
 532
 533 # CHECK:      Dispatch Width:    4
 534 # CHECK-NEXT: uOps Per Cycle:    0.99
 535 # CHECK-NEXT: IPC:               0.99
 536 # CHECK-NEXT: Block RThroughput: 4.0
 537
 538 # CHECK:      Instruction Info:
 539 # CHECK-NEXT: [1]: #uOps
 540 # CHECK-NEXT: [2]: Latency
 541 # CHECK-NEXT: [3]: RThroughput
 542 # CHECK-NEXT: [4]: MayLoad
 543 # CHECK-NEXT: [5]: MayStore
 544 # CHECK-NEXT: [6]: HasSideEffects (U)
 545
 546 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 547 # CHECK-NEXT:  1      1     1.00           *            movaps  %xmm0, (%rax)
 548 # CHECK-NEXT:  1      1     1.00           *            movaps  %xmm1, (%rcx)
 549 # CHECK-NEXT:  1      1     1.00           *            movaps  %xmm2, (%rdx)
 550 # CHECK-NEXT:  1      1     1.00           *            movaps  %xmm3, (%rbx)
 551
 552 # CHECK:      Dynamic Dispatch Stall Cycles:
 553 # CHECK-NEXT: RAT     - Register unavailable:                      0
 554 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 555 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            329  (81.6%)
 556 # CHECK-NEXT: LQ      - Load queue full:                           0
 557 # CHECK-NEXT: SQ      - Store queue full:                          0
 558 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 559
 560 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 561 # CHECK-NEXT: [# dispatched], [# cycles]
 562 # CHECK-NEXT:  0,              56  (13.9%)
 563 # CHECK-NEXT:  1,              329  (81.6%)
 564 # CHECK-NEXT:  3,              1  (0.2%)
 565 # CHECK-NEXT:  4,              17  (4.2%)
 566
 567 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 568 # CHECK-NEXT: [# issued], [# cycles]
 569 # CHECK-NEXT:  0,          3  (0.7%)
 570 # CHECK-NEXT:  1,          400  (99.3%)
 571
 572 # CHECK:      Scheduler's queue usage:
 573 # CHECK-NEXT: [1] Resource name.
 574 # CHECK-NEXT: [2] Average number of used buffer entries.
 575 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 576 # CHECK-NEXT: [4] Total number of buffer entries.
 577
 578 # CHECK:       [1]            [2]        [3]        [4]
 579 # CHECK-NEXT: SBPortAny        49         54         54
 580
 581 # CHECK:      Resources:
 582 # CHECK-NEXT: [0]   - SBDivider
 583 # CHECK-NEXT: [1]   - SBFPDivider
 584 # CHECK-NEXT: [2]   - SBPort0
 585 # CHECK-NEXT: [3]   - SBPort1
 586 # CHECK-NEXT: [4]   - SBPort4
 587 # CHECK-NEXT: [5]   - SBPort5
 588 # CHECK-NEXT: [6.0] - SBPort23
 589 # CHECK-NEXT: [6.1] - SBPort23
 590
 591 # CHECK:      Resource pressure per iteration:
 592 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 593 # CHECK-NEXT:  -      -      -      -     4.00    -     2.00   2.00
 594
 595 # CHECK:      Resource pressure by instruction:
 596 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 597 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movaps    %xmm0, (%rax)
 598 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movaps    %xmm1, (%rcx)
 599 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movaps    %xmm2, (%rdx)
 600 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movaps    %xmm3, (%rbx)
 601
 602 # CHECK:      Timeline view:
 603 # CHECK-NEXT: Index     0123456
 604
 605 # CHECK:      [0,0]     DeER ..   movaps        %xmm0, (%rax)
 606 # CHECK-NEXT: [0,1]     D=eER..   movaps        %xmm1, (%rcx)
 607 # CHECK-NEXT: [0,2]     D==eER.   movaps        %xmm2, (%rdx)
 608 # CHECK-NEXT: [0,3]     D===eER   movaps        %xmm3, (%rbx)
 609
 610 # CHECK:      Average Wait times (based on the timeline view):
 611 # CHECK-NEXT: [0]: Executions
 612 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 613 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 614 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 615
 616 # CHECK:            [0]    [1]    [2]    [3]
 617 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movaps       %xmm0, (%rax)
 618 # CHECK-NEXT: 1.     1     2.0    0.0    0.0       movaps       %xmm1, (%rcx)
 619 # CHECK-NEXT: 2.     1     3.0    0.0    0.0       movaps       %xmm2, (%rdx)
 620 # CHECK-NEXT: 3.     1     4.0    0.0    0.0       movaps       %xmm3, (%rbx)