llvm/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 # CHECK-NEXT: USH     - Uncategorised Structural Hazard:           0
  80
  81 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
  82 # CHECK-NEXT: [# dispatched], [# cycles]
  83 # CHECK-NEXT:  0,              56  (13.9%)
  84 # CHECK-NEXT:  1,              329  (81.6%)
  85 # CHECK-NEXT:  3,              1  (0.2%)
  86 # CHECK-NEXT:  4,              17  (4.2%)
  87
  88 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
  89 # CHECK-NEXT: [# issued], [# cycles]
  90 # CHECK-NEXT:  0,          3  (0.7%)
  91 # CHECK-NEXT:  1,          400  (99.3%)
  92
  93 # CHECK:      Scheduler's queue usage:
  94 # CHECK-NEXT: [1] Resource name.
  95 # CHECK-NEXT: [2] Average number of used buffer entries.
  96 # CHECK-NEXT: [3] Maximum number of used buffer entries.
  97 # CHECK-NEXT: [4] Total number of buffer entries.
  98
  99 # CHECK:       [1]            [2]        [3]        [4]
 100 # CHECK-NEXT: SBPortAny        49         54         54
 101
 102 # CHECK:      Resources:
 103 # CHECK-NEXT: [0]   - SBDivider
 104 # CHECK-NEXT: [1]   - SBFPDivider
 105 # CHECK-NEXT: [2]   - SBPort0
 106 # CHECK-NEXT: [3]   - SBPort1
 107 # CHECK-NEXT: [4]   - SBPort4
 108 # CHECK-NEXT: [5]   - SBPort5
 109 # CHECK-NEXT: [6.0] - SBPort23
 110 # CHECK-NEXT: [6.1] - SBPort23
 111
 112 # CHECK:      Resource pressure per iteration:
 113 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 114 # CHECK-NEXT:  -      -      -      -     4.00    -     2.00   2.00
 115
 116 # CHECK:      Resource pressure by instruction:
 117 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 118 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movb      %spl, (%rax)
 119 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movb      %bpl, (%rcx)
 120 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movb      %sil, (%rdx)
 121 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movb      %dil, (%rbx)
 122
 123 # CHECK:      Timeline view:
 124 # CHECK-NEXT: Index     0123456
 125
 126 # CHECK:      [0,0]     DeER ..   movb  %spl, (%rax)
 127 # CHECK-NEXT: [0,1]     D=eER..   movb  %bpl, (%rcx)
 128 # CHECK-NEXT: [0,2]     D==eER.   movb  %sil, (%rdx)
 129 # CHECK-NEXT: [0,3]     D===eER   movb  %dil, (%rbx)
 130
 131 # CHECK:      Average Wait times (based on the timeline view):
 132 # CHECK-NEXT: [0]: Executions
 133 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 134 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 135 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 136
 137 # CHECK:            [0]    [1]    [2]    [3]
 138 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movb %spl, (%rax)
 139 # CHECK-NEXT: 1.     1     2.0    1.0    0.0       movb %bpl, (%rcx)
 140 # CHECK-NEXT: 2.     1     3.0    1.0    0.0       movb %sil, (%rdx)
 141 # CHECK-NEXT: 3.     1     4.0    1.0    0.0       movb %dil, (%rbx)
 142 # CHECK-NEXT:        1     2.5    1.0    0.0       <total>
 143
 144 # CHECK:      [1] Code Region
 145
 146 # CHECK:      Iterations:        100
 147 # CHECK-NEXT: Instructions:      400
 148 # CHECK-NEXT: Total Cycles:      403
 149 # CHECK-NEXT: Total uOps:        400
 150
 151 # CHECK:      Dispatch Width:    4
 152 # CHECK-NEXT: uOps Per Cycle:    0.99
 153 # CHECK-NEXT: IPC:               0.99
 154 # CHECK-NEXT: Block RThroughput: 4.0
 155
 156 # CHECK:      Instruction Info:
 157 # CHECK-NEXT: [1]: #uOps
 158 # CHECK-NEXT: [2]: Latency
 159 # CHECK-NEXT: [3]: RThroughput
 160 # CHECK-NEXT: [4]: MayLoad
 161 # CHECK-NEXT: [5]: MayStore
 162 # CHECK-NEXT: [6]: HasSideEffects (U)
 163
 164 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 165 # CHECK-NEXT:  1      1     1.00           *            movw    %sp, (%rax)
 166 # CHECK-NEXT:  1      1     1.00           *            movw    %bp, (%rcx)
 167 # CHECK-NEXT:  1      1     1.00           *            movw    %si, (%rdx)
 168 # CHECK-NEXT:  1      1     1.00           *            movw    %di, (%rbx)
 169
 170 # CHECK:      Dynamic Dispatch Stall Cycles:
 171 # CHECK-NEXT: RAT     - Register unavailable:                      0
 172 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 173 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            329  (81.6%)
 174 # CHECK-NEXT: LQ      - Load queue full:                           0
 175 # CHECK-NEXT: SQ      - Store queue full:                          0
 176 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 177 # CHECK-NEXT: USH     - Uncategorised Structural Hazard:           0
 178
 179 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 180 # CHECK-NEXT: [# dispatched], [# cycles]
 181 # CHECK-NEXT:  0,              56  (13.9%)
 182 # CHECK-NEXT:  1,              329  (81.6%)
 183 # CHECK-NEXT:  3,              1  (0.2%)
 184 # CHECK-NEXT:  4,              17  (4.2%)
 185
 186 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 187 # CHECK-NEXT: [# issued], [# cycles]
 188 # CHECK-NEXT:  0,          3  (0.7%)
 189 # CHECK-NEXT:  1,          400  (99.3%)
 190
 191 # CHECK:      Scheduler's queue usage:
 192 # CHECK-NEXT: [1] Resource name.
 193 # CHECK-NEXT: [2] Average number of used buffer entries.
 194 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 195 # CHECK-NEXT: [4] Total number of buffer entries.
 196
 197 # CHECK:       [1]            [2]        [3]        [4]
 198 # CHECK-NEXT: SBPortAny        49         54         54
 199
 200 # CHECK:      Resources:
 201 # CHECK-NEXT: [0]   - SBDivider
 202 # CHECK-NEXT: [1]   - SBFPDivider
 203 # CHECK-NEXT: [2]   - SBPort0
 204 # CHECK-NEXT: [3]   - SBPort1
 205 # CHECK-NEXT: [4]   - SBPort4
 206 # CHECK-NEXT: [5]   - SBPort5
 207 # CHECK-NEXT: [6.0] - SBPort23
 208 # CHECK-NEXT: [6.1] - SBPort23
 209
 210 # CHECK:      Resource pressure per iteration:
 211 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 212 # CHECK-NEXT:  -      -      -      -     4.00    -     2.00   2.00
 213
 214 # CHECK:      Resource pressure by instruction:
 215 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 216 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movw      %sp, (%rax)
 217 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movw      %bp, (%rcx)
 218 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movw      %si, (%rdx)
 219 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movw      %di, (%rbx)
 220
 221 # CHECK:      Timeline view:
 222 # CHECK-NEXT: Index     0123456
 223
 224 # CHECK:      [0,0]     DeER ..   movw  %sp, (%rax)
 225 # CHECK-NEXT: [0,1]     D=eER..   movw  %bp, (%rcx)
 226 # CHECK-NEXT: [0,2]     D==eER.   movw  %si, (%rdx)
 227 # CHECK-NEXT: [0,3]     D===eER   movw  %di, (%rbx)
 228
 229 # CHECK:      Average Wait times (based on the timeline view):
 230 # CHECK-NEXT: [0]: Executions
 231 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 232 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 233 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 234
 235 # CHECK:            [0]    [1]    [2]    [3]
 236 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movw %sp, (%rax)
 237 # CHECK-NEXT: 1.     1     2.0    1.0    0.0       movw %bp, (%rcx)
 238 # CHECK-NEXT: 2.     1     3.0    1.0    0.0       movw %si, (%rdx)
 239 # CHECK-NEXT: 3.     1     4.0    1.0    0.0       movw %di, (%rbx)
 240 # CHECK-NEXT:        1     2.5    1.0    0.0       <total>
 241
 242 # CHECK:      [2] Code Region
 243
 244 # CHECK:      Iterations:        100
 245 # CHECK-NEXT: Instructions:      400
 246 # CHECK-NEXT: Total Cycles:      403
 247 # CHECK-NEXT: Total uOps:        400
 248
 249 # CHECK:      Dispatch Width:    4
 250 # CHECK-NEXT: uOps Per Cycle:    0.99
 251 # CHECK-NEXT: IPC:               0.99
 252 # CHECK-NEXT: Block RThroughput: 4.0
 253
 254 # CHECK:      Instruction Info:
 255 # CHECK-NEXT: [1]: #uOps
 256 # CHECK-NEXT: [2]: Latency
 257 # CHECK-NEXT: [3]: RThroughput
 258 # CHECK-NEXT: [4]: MayLoad
 259 # CHECK-NEXT: [5]: MayStore
 260 # CHECK-NEXT: [6]: HasSideEffects (U)
 261
 262 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 263 # CHECK-NEXT:  1      1     1.00           *            movl    %esp, (%rax)
 264 # CHECK-NEXT:  1      1     1.00           *            movl    %ebp, (%rcx)
 265 # CHECK-NEXT:  1      1     1.00           *            movl    %esi, (%rdx)
 266 # CHECK-NEXT:  1      1     1.00           *            movl    %edi, (%rbx)
 267
 268 # CHECK:      Dynamic Dispatch Stall Cycles:
 269 # CHECK-NEXT: RAT     - Register unavailable:                      0
 270 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 271 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            329  (81.6%)
 272 # CHECK-NEXT: LQ      - Load queue full:                           0
 273 # CHECK-NEXT: SQ      - Store queue full:                          0
 274 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 275 # CHECK-NEXT: USH     - Uncategorised Structural Hazard:           0
 276
 277 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 278 # CHECK-NEXT: [# dispatched], [# cycles]
 279 # CHECK-NEXT:  0,              56  (13.9%)
 280 # CHECK-NEXT:  1,              329  (81.6%)
 281 # CHECK-NEXT:  3,              1  (0.2%)
 282 # CHECK-NEXT:  4,              17  (4.2%)
 283
 284 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 285 # CHECK-NEXT: [# issued], [# cycles]
 286 # CHECK-NEXT:  0,          3  (0.7%)
 287 # CHECK-NEXT:  1,          400  (99.3%)
 288
 289 # CHECK:      Scheduler's queue usage:
 290 # CHECK-NEXT: [1] Resource name.
 291 # CHECK-NEXT: [2] Average number of used buffer entries.
 292 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 293 # CHECK-NEXT: [4] Total number of buffer entries.
 294
 295 # CHECK:       [1]            [2]        [3]        [4]
 296 # CHECK-NEXT: SBPortAny        49         54         54
 297
 298 # CHECK:      Resources:
 299 # CHECK-NEXT: [0]   - SBDivider
 300 # CHECK-NEXT: [1]   - SBFPDivider
 301 # CHECK-NEXT: [2]   - SBPort0
 302 # CHECK-NEXT: [3]   - SBPort1
 303 # CHECK-NEXT: [4]   - SBPort4
 304 # CHECK-NEXT: [5]   - SBPort5
 305 # CHECK-NEXT: [6.0] - SBPort23
 306 # CHECK-NEXT: [6.1] - SBPort23
 307
 308 # CHECK:      Resource pressure per iteration:
 309 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 310 # CHECK-NEXT:  -      -      -      -     4.00    -     2.00   2.00
 311
 312 # CHECK:      Resource pressure by instruction:
 313 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 314 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movl      %esp, (%rax)
 315 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movl      %ebp, (%rcx)
 316 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movl      %esi, (%rdx)
 317 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movl      %edi, (%rbx)
 318
 319 # CHECK:      Timeline view:
 320 # CHECK-NEXT: Index     0123456
 321
 322 # CHECK:      [0,0]     DeER ..   movl  %esp, (%rax)
 323 # CHECK-NEXT: [0,1]     D=eER..   movl  %ebp, (%rcx)
 324 # CHECK-NEXT: [0,2]     D==eER.   movl  %esi, (%rdx)
 325 # CHECK-NEXT: [0,3]     D===eER   movl  %edi, (%rbx)
 326
 327 # CHECK:      Average Wait times (based on the timeline view):
 328 # CHECK-NEXT: [0]: Executions
 329 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 330 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 331 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 332
 333 # CHECK:            [0]    [1]    [2]    [3]
 334 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movl %esp, (%rax)
 335 # CHECK-NEXT: 1.     1     2.0    1.0    0.0       movl %ebp, (%rcx)
 336 # CHECK-NEXT: 2.     1     3.0    1.0    0.0       movl %esi, (%rdx)
 337 # CHECK-NEXT: 3.     1     4.0    1.0    0.0       movl %edi, (%rbx)
 338 # CHECK-NEXT:        1     2.5    1.0    0.0       <total>
 339
 340 # CHECK:      [3] Code Region
 341
 342 # CHECK:      Iterations:        100
 343 # CHECK-NEXT: Instructions:      400
 344 # CHECK-NEXT: Total Cycles:      403
 345 # CHECK-NEXT: Total uOps:        400
 346
 347 # CHECK:      Dispatch Width:    4
 348 # CHECK-NEXT: uOps Per Cycle:    0.99
 349 # CHECK-NEXT: IPC:               0.99
 350 # CHECK-NEXT: Block RThroughput: 4.0
 351
 352 # CHECK:      Instruction Info:
 353 # CHECK-NEXT: [1]: #uOps
 354 # CHECK-NEXT: [2]: Latency
 355 # CHECK-NEXT: [3]: RThroughput
 356 # CHECK-NEXT: [4]: MayLoad
 357 # CHECK-NEXT: [5]: MayStore
 358 # CHECK-NEXT: [6]: HasSideEffects (U)
 359
 360 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 361 # CHECK-NEXT:  1      1     1.00           *            movq    %rsp, (%rax)
 362 # CHECK-NEXT:  1      1     1.00           *            movq    %rbp, (%rcx)
 363 # CHECK-NEXT:  1      1     1.00           *            movq    %rsi, (%rdx)
 364 # CHECK-NEXT:  1      1     1.00           *            movq    %rdi, (%rbx)
 365
 366 # CHECK:      Dynamic Dispatch Stall Cycles:
 367 # CHECK-NEXT: RAT     - Register unavailable:                      0
 368 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 369 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            329  (81.6%)
 370 # CHECK-NEXT: LQ      - Load queue full:                           0
 371 # CHECK-NEXT: SQ      - Store queue full:                          0
 372 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 373 # CHECK-NEXT: USH     - Uncategorised Structural Hazard:           0
 374
 375 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 376 # CHECK-NEXT: [# dispatched], [# cycles]
 377 # CHECK-NEXT:  0,              56  (13.9%)
 378 # CHECK-NEXT:  1,              329  (81.6%)
 379 # CHECK-NEXT:  3,              1  (0.2%)
 380 # CHECK-NEXT:  4,              17  (4.2%)
 381
 382 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 383 # CHECK-NEXT: [# issued], [# cycles]
 384 # CHECK-NEXT:  0,          3  (0.7%)
 385 # CHECK-NEXT:  1,          400  (99.3%)
 386
 387 # CHECK:      Scheduler's queue usage:
 388 # CHECK-NEXT: [1] Resource name.
 389 # CHECK-NEXT: [2] Average number of used buffer entries.
 390 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 391 # CHECK-NEXT: [4] Total number of buffer entries.
 392
 393 # CHECK:       [1]            [2]        [3]        [4]
 394 # CHECK-NEXT: SBPortAny        49         54         54
 395
 396 # CHECK:      Resources:
 397 # CHECK-NEXT: [0]   - SBDivider
 398 # CHECK-NEXT: [1]   - SBFPDivider
 399 # CHECK-NEXT: [2]   - SBPort0
 400 # CHECK-NEXT: [3]   - SBPort1
 401 # CHECK-NEXT: [4]   - SBPort4
 402 # CHECK-NEXT: [5]   - SBPort5
 403 # CHECK-NEXT: [6.0] - SBPort23
 404 # CHECK-NEXT: [6.1] - SBPort23
 405
 406 # CHECK:      Resource pressure per iteration:
 407 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 408 # CHECK-NEXT:  -      -      -      -     4.00    -     2.00   2.00
 409
 410 # CHECK:      Resource pressure by instruction:
 411 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 412 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movq      %rsp, (%rax)
 413 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movq      %rbp, (%rcx)
 414 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movq      %rsi, (%rdx)
 415 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movq      %rdi, (%rbx)
 416
 417 # CHECK:      Timeline view:
 418 # CHECK-NEXT: Index     0123456
 419
 420 # CHECK:      [0,0]     DeER ..   movq  %rsp, (%rax)
 421 # CHECK-NEXT: [0,1]     D=eER..   movq  %rbp, (%rcx)
 422 # CHECK-NEXT: [0,2]     D==eER.   movq  %rsi, (%rdx)
 423 # CHECK-NEXT: [0,3]     D===eER   movq  %rdi, (%rbx)
 424
 425 # CHECK:      Average Wait times (based on the timeline view):
 426 # CHECK-NEXT: [0]: Executions
 427 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 428 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 429 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 430
 431 # CHECK:            [0]    [1]    [2]    [3]
 432 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movq %rsp, (%rax)
 433 # CHECK-NEXT: 1.     1     2.0    1.0    0.0       movq %rbp, (%rcx)
 434 # CHECK-NEXT: 2.     1     3.0    1.0    0.0       movq %rsi, (%rdx)
 435 # CHECK-NEXT: 3.     1     4.0    1.0    0.0       movq %rdi, (%rbx)
 436 # CHECK-NEXT:        1     2.5    1.0    0.0       <total>
 437
 438 # CHECK:      [4] Code Region
 439
 440 # CHECK:      Iterations:        100
 441 # CHECK-NEXT: Instructions:      400
 442 # CHECK-NEXT: Total Cycles:      403
 443 # CHECK-NEXT: Total uOps:        400
 444
 445 # CHECK:      Dispatch Width:    4
 446 # CHECK-NEXT: uOps Per Cycle:    0.99
 447 # CHECK-NEXT: IPC:               0.99
 448 # CHECK-NEXT: Block RThroughput: 4.0
 449
 450 # CHECK:      Instruction Info:
 451 # CHECK-NEXT: [1]: #uOps
 452 # CHECK-NEXT: [2]: Latency
 453 # CHECK-NEXT: [3]: RThroughput
 454 # CHECK-NEXT: [4]: MayLoad
 455 # CHECK-NEXT: [5]: MayStore
 456 # CHECK-NEXT: [6]: HasSideEffects (U)
 457
 458 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 459 # CHECK-NEXT:  1      1     1.00           *      U     movd    %mm0, (%rax)
 460 # CHECK-NEXT:  1      1     1.00           *      U     movd    %mm1, (%rcx)
 461 # CHECK-NEXT:  1      1     1.00           *      U     movd    %mm2, (%rdx)
 462 # CHECK-NEXT:  1      1     1.00           *      U     movd    %mm3, (%rbx)
 463
 464 # CHECK:      Dynamic Dispatch Stall Cycles:
 465 # CHECK-NEXT: RAT     - Register unavailable:                      0
 466 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 467 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            329  (81.6%)
 468 # CHECK-NEXT: LQ      - Load queue full:                           0
 469 # CHECK-NEXT: SQ      - Store queue full:                          0
 470 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 471 # CHECK-NEXT: USH     - Uncategorised Structural Hazard:           0
 472
 473 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 474 # CHECK-NEXT: [# dispatched], [# cycles]
 475 # CHECK-NEXT:  0,              56  (13.9%)
 476 # CHECK-NEXT:  1,              329  (81.6%)
 477 # CHECK-NEXT:  3,              1  (0.2%)
 478 # CHECK-NEXT:  4,              17  (4.2%)
 479
 480 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 481 # CHECK-NEXT: [# issued], [# cycles]
 482 # CHECK-NEXT:  0,          3  (0.7%)
 483 # CHECK-NEXT:  1,          400  (99.3%)
 484
 485 # CHECK:      Scheduler's queue usage:
 486 # CHECK-NEXT: [1] Resource name.
 487 # CHECK-NEXT: [2] Average number of used buffer entries.
 488 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 489 # CHECK-NEXT: [4] Total number of buffer entries.
 490
 491 # CHECK:       [1]            [2]        [3]        [4]
 492 # CHECK-NEXT: SBPortAny        49         54         54
 493
 494 # CHECK:      Resources:
 495 # CHECK-NEXT: [0]   - SBDivider
 496 # CHECK-NEXT: [1]   - SBFPDivider
 497 # CHECK-NEXT: [2]   - SBPort0
 498 # CHECK-NEXT: [3]   - SBPort1
 499 # CHECK-NEXT: [4]   - SBPort4
 500 # CHECK-NEXT: [5]   - SBPort5
 501 # CHECK-NEXT: [6.0] - SBPort23
 502 # CHECK-NEXT: [6.1] - SBPort23
 503
 504 # CHECK:      Resource pressure per iteration:
 505 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 506 # CHECK-NEXT:  -      -      -      -     4.00    -     2.00   2.00
 507
 508 # CHECK:      Resource pressure by instruction:
 509 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 510 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movd      %mm0, (%rax)
 511 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movd      %mm1, (%rcx)
 512 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movd      %mm2, (%rdx)
 513 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movd      %mm3, (%rbx)
 514
 515 # CHECK:      Timeline view:
 516 # CHECK-NEXT: Index     0123456
 517
 518 # CHECK:      [0,0]     DeER ..   movd  %mm0, (%rax)
 519 # CHECK-NEXT: [0,1]     D=eER..   movd  %mm1, (%rcx)
 520 # CHECK-NEXT: [0,2]     D==eER.   movd  %mm2, (%rdx)
 521 # CHECK-NEXT: [0,3]     D===eER   movd  %mm3, (%rbx)
 522
 523 # CHECK:      Average Wait times (based on the timeline view):
 524 # CHECK-NEXT: [0]: Executions
 525 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 526 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 527 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 528
 529 # CHECK:            [0]    [1]    [2]    [3]
 530 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movd %mm0, (%rax)
 531 # CHECK-NEXT: 1.     1     2.0    1.0    0.0       movd %mm1, (%rcx)
 532 # CHECK-NEXT: 2.     1     3.0    1.0    0.0       movd %mm2, (%rdx)
 533 # CHECK-NEXT: 3.     1     4.0    1.0    0.0       movd %mm3, (%rbx)
 534 # CHECK-NEXT:        1     2.5    1.0    0.0       <total>
 535
 536 # CHECK:      [5] Code Region
 537
 538 # CHECK:      Iterations:        100
 539 # CHECK-NEXT: Instructions:      400
 540 # CHECK-NEXT: Total Cycles:      403
 541 # CHECK-NEXT: Total uOps:        400
 542
 543 # CHECK:      Dispatch Width:    4
 544 # CHECK-NEXT: uOps Per Cycle:    0.99
 545 # CHECK-NEXT: IPC:               0.99
 546 # CHECK-NEXT: Block RThroughput: 4.0
 547
 548 # CHECK:      Instruction Info:
 549 # CHECK-NEXT: [1]: #uOps
 550 # CHECK-NEXT: [2]: Latency
 551 # CHECK-NEXT: [3]: RThroughput
 552 # CHECK-NEXT: [4]: MayLoad
 553 # CHECK-NEXT: [5]: MayStore
 554 # CHECK-NEXT: [6]: HasSideEffects (U)
 555
 556 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 557 # CHECK-NEXT:  1      1     1.00           *            movaps  %xmm0, (%rax)
 558 # CHECK-NEXT:  1      1     1.00           *            movaps  %xmm1, (%rcx)
 559 # CHECK-NEXT:  1      1     1.00           *            movaps  %xmm2, (%rdx)
 560 # CHECK-NEXT:  1      1     1.00           *            movaps  %xmm3, (%rbx)
 561
 562 # CHECK:      Dynamic Dispatch Stall Cycles:
 563 # CHECK-NEXT: RAT     - Register unavailable:                      0
 564 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 565 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            329  (81.6%)
 566 # CHECK-NEXT: LQ      - Load queue full:                           0
 567 # CHECK-NEXT: SQ      - Store queue full:                          0
 568 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 569 # CHECK-NEXT: USH     - Uncategorised Structural Hazard:           0
 570
 571 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 572 # CHECK-NEXT: [# dispatched], [# cycles]
 573 # CHECK-NEXT:  0,              56  (13.9%)
 574 # CHECK-NEXT:  1,              329  (81.6%)
 575 # CHECK-NEXT:  3,              1  (0.2%)
 576 # CHECK-NEXT:  4,              17  (4.2%)
 577
 578 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 579 # CHECK-NEXT: [# issued], [# cycles]
 580 # CHECK-NEXT:  0,          3  (0.7%)
 581 # CHECK-NEXT:  1,          400  (99.3%)
 582
 583 # CHECK:      Scheduler's queue usage:
 584 # CHECK-NEXT: [1] Resource name.
 585 # CHECK-NEXT: [2] Average number of used buffer entries.
 586 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 587 # CHECK-NEXT: [4] Total number of buffer entries.
 588
 589 # CHECK:       [1]            [2]        [3]        [4]
 590 # CHECK-NEXT: SBPortAny        49         54         54
 591
 592 # CHECK:      Resources:
 593 # CHECK-NEXT: [0]   - SBDivider
 594 # CHECK-NEXT: [1]   - SBFPDivider
 595 # CHECK-NEXT: [2]   - SBPort0
 596 # CHECK-NEXT: [3]   - SBPort1
 597 # CHECK-NEXT: [4]   - SBPort4
 598 # CHECK-NEXT: [5]   - SBPort5
 599 # CHECK-NEXT: [6.0] - SBPort23
 600 # CHECK-NEXT: [6.1] - SBPort23
 601
 602 # CHECK:      Resource pressure per iteration:
 603 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 604 # CHECK-NEXT:  -      -      -      -     4.00    -     2.00   2.00
 605
 606 # CHECK:      Resource pressure by instruction:
 607 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 608 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movaps    %xmm0, (%rax)
 609 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movaps    %xmm1, (%rcx)
 610 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movaps    %xmm2, (%rdx)
 611 # CHECK-NEXT:  -      -      -      -     1.00    -     1.00    -     movaps    %xmm3, (%rbx)
 612
 613 # CHECK:      Timeline view:
 614 # CHECK-NEXT: Index     0123456
 615
 616 # CHECK:      [0,0]     DeER ..   movaps        %xmm0, (%rax)
 617 # CHECK-NEXT: [0,1]     D=eER..   movaps        %xmm1, (%rcx)
 618 # CHECK-NEXT: [0,2]     D==eER.   movaps        %xmm2, (%rdx)
 619 # CHECK-NEXT: [0,3]     D===eER   movaps        %xmm3, (%rbx)
 620
 621 # CHECK:      Average Wait times (based on the timeline view):
 622 # CHECK-NEXT: [0]: Executions
 623 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 624 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 625 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 626
 627 # CHECK:            [0]    [1]    [2]    [3]
 628 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movaps       %xmm0, (%rax)
 629 # CHECK-NEXT: 1.     1     2.0    1.0    0.0       movaps       %xmm1, (%rcx)
 630 # CHECK-NEXT: 2.     1     3.0    1.0    0.0       movaps       %xmm2, (%rdx)
 631 # CHECK-NEXT: 3.     1     4.0    1.0    0.0       movaps       %xmm3, (%rbx)
 632 # CHECK-NEXT:        1     2.5    1.0    0.0       <total>