test/tools/llvm-mca/X86/Barcelona/load-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 < %s | FileCheck %s
   3
   4 # LLVM-MCA-BEGIN
   5 movb %spl, (%rax)
   6 movb (%rcx), %bpl
   7 movb (%rdx), %sil
   8 movb %dil, (%rbx)
   9 # LLVM-MCA-END
  10
  11 # LLVM-MCA-BEGIN
  12 movw %sp, (%rax)
  13 movw (%rcx), %bp
  14 movw (%rdx), %si
  15 movw %di, (%rbx)
  16 # LLVM-MCA-END
  17
  18 # LLVM-MCA-BEGIN
  19 movl %esp, (%rax)
  20 movl (%rcx), %ebp
  21 movl (%rdx), %esi
  22 movl %edi, (%rbx)
  23 # LLVM-MCA-END
  24
  25 # LLVM-MCA-BEGIN
  26 movq %rsp, (%rax)
  27 movq (%rcx), %rbp
  28 movq (%rdx), %rsi
  29 movq %rdi, (%rbx)
  30 # LLVM-MCA-END
  31
  32 # LLVM-MCA-BEGIN
  33 movd %mm0, (%rax)
  34 movd (%rcx), %mm1
  35 movd (%rdx), %mm2
  36 movd %mm3, (%rbx)
  37 # LLVM-MCA-END
  38
  39 # LLVM-MCA-BEGIN
  40 movaps %xmm0, (%rax)
  41 movaps (%rcx), %xmm1
  42 movaps (%rdx), %xmm2
  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:      208
  51 # CHECK-NEXT: Total uOps:        400
  52
  53 # CHECK:      Dispatch Width:    4
  54 # CHECK-NEXT: uOps Per Cycle:    1.92
  55 # CHECK-NEXT: IPC:               1.92
  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      1     1.00           *            movb    %spl, (%rax)
  68 # CHECK-NEXT:  1      5     0.50    *                   movb    (%rcx), %bpl
  69 # CHECK-NEXT:  1      5     0.50    *                   movb    (%rdx), %sil
  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:                            147  (70.7%)
  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,              34  (16.3%)
  83 # CHECK-NEXT:  2,              148  (71.2%)
  84 # CHECK-NEXT:  4,              26  (12.5%)
  85
  86 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
  87 # CHECK-NEXT: [# issued], [# cycles]
  88 # CHECK-NEXT:  0,          3  (1.4%)
  89 # CHECK-NEXT:  1,          10  (4.8%)
  90 # CHECK-NEXT:  2,          195  (93.8%)
  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        45         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:  -      -      -      -     2.00    -     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    -     movb      (%rcx), %bpl
 119 # CHECK-NEXT:  -      -      -      -      -      -     0.95   0.05   movb      (%rdx), %sil
 120 # CHECK-NEXT:  -      -      -      -     1.00    -     0.05   0.95   movb      %dil, (%rbx)
 121
 122 # CHECK:      Timeline view:
 123 # CHECK-NEXT: Index     0123456789
 124
 125 # CHECK:      [0,0]     DeER .   .   movb       %spl, (%rax)
 126 # CHECK-NEXT: [0,1]     DeeeeeER .   movb       (%rcx), %bpl
 127 # CHECK-NEXT: [0,2]     D=eeeeeER.   movb       (%rdx), %sil
 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     1.0    1.0    0.0       movb (%rcx), %bpl
 139 # CHECK-NEXT: 2.     1     2.0    2.0    0.0       movb (%rdx), %sil
 140 # CHECK-NEXT: 3.     1     7.0    0.0    0.0       movb %dil, (%rbx)
 141 # CHECK-NEXT:        1     2.8    1.0    0.0       <total>
 142
 143 # CHECK:      [1] Code Region
 144
 145 # CHECK:      Iterations:        100
 146 # CHECK-NEXT: Instructions:      400
 147 # CHECK-NEXT: Total Cycles:      208
 148 # CHECK-NEXT: Total uOps:        400
 149
 150 # CHECK:      Dispatch Width:    4
 151 # CHECK-NEXT: uOps Per Cycle:    1.92
 152 # CHECK-NEXT: IPC:               1.92
 153 # CHECK-NEXT: Block RThroughput: 2.0
 154
 155 # CHECK:      Instruction Info:
 156 # CHECK-NEXT: [1]: #uOps
 157 # CHECK-NEXT: [2]: Latency
 158 # CHECK-NEXT: [3]: RThroughput
 159 # CHECK-NEXT: [4]: MayLoad
 160 # CHECK-NEXT: [5]: MayStore
 161 # CHECK-NEXT: [6]: HasSideEffects (U)
 162
 163 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 164 # CHECK-NEXT:  1      1     1.00           *            movw    %sp, (%rax)
 165 # CHECK-NEXT:  1      5     0.50    *                   movw    (%rcx), %bp
 166 # CHECK-NEXT:  1      5     0.50    *                   movw    (%rdx), %si
 167 # CHECK-NEXT:  1      1     1.00           *            movw    %di, (%rbx)
 168
 169 # CHECK:      Dynamic Dispatch Stall Cycles:
 170 # CHECK-NEXT: RAT     - Register unavailable:                      0
 171 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 172 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            147  (70.7%)
 173 # CHECK-NEXT: LQ      - Load queue full:                           0
 174 # CHECK-NEXT: SQ      - Store queue full:                          0
 175 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 176
 177 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 178 # CHECK-NEXT: [# dispatched], [# cycles]
 179 # CHECK-NEXT:  0,              34  (16.3%)
 180 # CHECK-NEXT:  2,              148  (71.2%)
 181 # CHECK-NEXT:  4,              26  (12.5%)
 182
 183 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 184 # CHECK-NEXT: [# issued], [# cycles]
 185 # CHECK-NEXT:  0,          3  (1.4%)
 186 # CHECK-NEXT:  1,          10  (4.8%)
 187 # CHECK-NEXT:  2,          195  (93.8%)
 188
 189 # CHECK:      Scheduler's queue usage:
 190 # CHECK-NEXT: [1] Resource name.
 191 # CHECK-NEXT: [2] Average number of used buffer entries.
 192 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 193 # CHECK-NEXT: [4] Total number of buffer entries.
 194
 195 # CHECK:       [1]            [2]        [3]        [4]
 196 # CHECK-NEXT: SBPortAny        45         54         54
 197
 198 # CHECK:      Resources:
 199 # CHECK-NEXT: [0]   - SBDivider
 200 # CHECK-NEXT: [1]   - SBFPDivider
 201 # CHECK-NEXT: [2]   - SBPort0
 202 # CHECK-NEXT: [3]   - SBPort1
 203 # CHECK-NEXT: [4]   - SBPort4
 204 # CHECK-NEXT: [5]   - SBPort5
 205 # CHECK-NEXT: [6.0] - SBPort23
 206 # CHECK-NEXT: [6.1] - SBPort23
 207
 208 # CHECK:      Resource pressure per iteration:
 209 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 210 # CHECK-NEXT:  -      -      -      -     2.00    -     2.00   2.00
 211
 212 # CHECK:      Resource pressure by instruction:
 213 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 214 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movw      %sp, (%rax)
 215 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movw      (%rcx), %bp
 216 # CHECK-NEXT:  -      -      -      -      -      -     0.95   0.05   movw      (%rdx), %si
 217 # CHECK-NEXT:  -      -      -      -     1.00    -     0.05   0.95   movw      %di, (%rbx)
 218
 219 # CHECK:      Timeline view:
 220 # CHECK-NEXT: Index     0123456789
 221
 222 # CHECK:      [0,0]     DeER .   .   movw       %sp, (%rax)
 223 # CHECK-NEXT: [0,1]     DeeeeeER .   movw       (%rcx), %bp
 224 # CHECK-NEXT: [0,2]     D=eeeeeER.   movw       (%rdx), %si
 225 # CHECK-NEXT: [0,3]     D======eER   movw       %di, (%rbx)
 226
 227 # CHECK:      Average Wait times (based on the timeline view):
 228 # CHECK-NEXT: [0]: Executions
 229 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 230 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 231 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 232
 233 # CHECK:            [0]    [1]    [2]    [3]
 234 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movw %sp, (%rax)
 235 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movw (%rcx), %bp
 236 # CHECK-NEXT: 2.     1     2.0    2.0    0.0       movw (%rdx), %si
 237 # CHECK-NEXT: 3.     1     7.0    0.0    0.0       movw %di, (%rbx)
 238 # CHECK-NEXT:        1     2.8    1.0    0.0       <total>
 239
 240 # CHECK:      [2] Code Region
 241
 242 # CHECK:      Iterations:        100
 243 # CHECK-NEXT: Instructions:      400
 244 # CHECK-NEXT: Total Cycles:      208
 245 # CHECK-NEXT: Total uOps:        400
 246
 247 # CHECK:      Dispatch Width:    4
 248 # CHECK-NEXT: uOps Per Cycle:    1.92
 249 # CHECK-NEXT: IPC:               1.92
 250 # CHECK-NEXT: Block RThroughput: 2.0
 251
 252 # CHECK:      Instruction Info:
 253 # CHECK-NEXT: [1]: #uOps
 254 # CHECK-NEXT: [2]: Latency
 255 # CHECK-NEXT: [3]: RThroughput
 256 # CHECK-NEXT: [4]: MayLoad
 257 # CHECK-NEXT: [5]: MayStore
 258 # CHECK-NEXT: [6]: HasSideEffects (U)
 259
 260 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 261 # CHECK-NEXT:  1      1     1.00           *            movl    %esp, (%rax)
 262 # CHECK-NEXT:  1      5     0.50    *                   movl    (%rcx), %ebp
 263 # CHECK-NEXT:  1      5     0.50    *                   movl    (%rdx), %esi
 264 # CHECK-NEXT:  1      1     1.00           *            movl    %edi, (%rbx)
 265
 266 # CHECK:      Dynamic Dispatch Stall Cycles:
 267 # CHECK-NEXT: RAT     - Register unavailable:                      0
 268 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 269 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            147  (70.7%)
 270 # CHECK-NEXT: LQ      - Load queue full:                           0
 271 # CHECK-NEXT: SQ      - Store queue full:                          0
 272 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 273
 274 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 275 # CHECK-NEXT: [# dispatched], [# cycles]
 276 # CHECK-NEXT:  0,              34  (16.3%)
 277 # CHECK-NEXT:  2,              148  (71.2%)
 278 # CHECK-NEXT:  4,              26  (12.5%)
 279
 280 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 281 # CHECK-NEXT: [# issued], [# cycles]
 282 # CHECK-NEXT:  0,          3  (1.4%)
 283 # CHECK-NEXT:  1,          10  (4.8%)
 284 # CHECK-NEXT:  2,          195  (93.8%)
 285
 286 # CHECK:      Scheduler's queue usage:
 287 # CHECK-NEXT: [1] Resource name.
 288 # CHECK-NEXT: [2] Average number of used buffer entries.
 289 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 290 # CHECK-NEXT: [4] Total number of buffer entries.
 291
 292 # CHECK:       [1]            [2]        [3]        [4]
 293 # CHECK-NEXT: SBPortAny        45         54         54
 294
 295 # CHECK:      Resources:
 296 # CHECK-NEXT: [0]   - SBDivider
 297 # CHECK-NEXT: [1]   - SBFPDivider
 298 # CHECK-NEXT: [2]   - SBPort0
 299 # CHECK-NEXT: [3]   - SBPort1
 300 # CHECK-NEXT: [4]   - SBPort4
 301 # CHECK-NEXT: [5]   - SBPort5
 302 # CHECK-NEXT: [6.0] - SBPort23
 303 # CHECK-NEXT: [6.1] - SBPort23
 304
 305 # CHECK:      Resource pressure per iteration:
 306 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 307 # CHECK-NEXT:  -      -      -      -     2.00    -     2.00   2.00
 308
 309 # CHECK:      Resource pressure by instruction:
 310 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 311 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movl      %esp, (%rax)
 312 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movl      (%rcx), %ebp
 313 # CHECK-NEXT:  -      -      -      -      -      -     0.95   0.05   movl      (%rdx), %esi
 314 # CHECK-NEXT:  -      -      -      -     1.00    -     0.05   0.95   movl      %edi, (%rbx)
 315
 316 # CHECK:      Timeline view:
 317 # CHECK-NEXT: Index     0123456789
 318
 319 # CHECK:      [0,0]     DeER .   .   movl       %esp, (%rax)
 320 # CHECK-NEXT: [0,1]     DeeeeeER .   movl       (%rcx), %ebp
 321 # CHECK-NEXT: [0,2]     D=eeeeeER.   movl       (%rdx), %esi
 322 # CHECK-NEXT: [0,3]     D======eER   movl       %edi, (%rbx)
 323
 324 # CHECK:      Average Wait times (based on the timeline view):
 325 # CHECK-NEXT: [0]: Executions
 326 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 327 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 328 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 329
 330 # CHECK:            [0]    [1]    [2]    [3]
 331 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movl %esp, (%rax)
 332 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movl (%rcx), %ebp
 333 # CHECK-NEXT: 2.     1     2.0    2.0    0.0       movl (%rdx), %esi
 334 # CHECK-NEXT: 3.     1     7.0    0.0    0.0       movl %edi, (%rbx)
 335 # CHECK-NEXT:        1     2.8    1.0    0.0       <total>
 336
 337 # CHECK:      [3] Code Region
 338
 339 # CHECK:      Iterations:        100
 340 # CHECK-NEXT: Instructions:      400
 341 # CHECK-NEXT: Total Cycles:      208
 342 # CHECK-NEXT: Total uOps:        400
 343
 344 # CHECK:      Dispatch Width:    4
 345 # CHECK-NEXT: uOps Per Cycle:    1.92
 346 # CHECK-NEXT: IPC:               1.92
 347 # CHECK-NEXT: Block RThroughput: 2.0
 348
 349 # CHECK:      Instruction Info:
 350 # CHECK-NEXT: [1]: #uOps
 351 # CHECK-NEXT: [2]: Latency
 352 # CHECK-NEXT: [3]: RThroughput
 353 # CHECK-NEXT: [4]: MayLoad
 354 # CHECK-NEXT: [5]: MayStore
 355 # CHECK-NEXT: [6]: HasSideEffects (U)
 356
 357 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 358 # CHECK-NEXT:  1      1     1.00           *            movq    %rsp, (%rax)
 359 # CHECK-NEXT:  1      5     0.50    *                   movq    (%rcx), %rbp
 360 # CHECK-NEXT:  1      5     0.50    *                   movq    (%rdx), %rsi
 361 # CHECK-NEXT:  1      1     1.00           *            movq    %rdi, (%rbx)
 362
 363 # CHECK:      Dynamic Dispatch Stall Cycles:
 364 # CHECK-NEXT: RAT     - Register unavailable:                      0
 365 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 366 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            147  (70.7%)
 367 # CHECK-NEXT: LQ      - Load queue full:                           0
 368 # CHECK-NEXT: SQ      - Store queue full:                          0
 369 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 370
 371 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 372 # CHECK-NEXT: [# dispatched], [# cycles]
 373 # CHECK-NEXT:  0,              34  (16.3%)
 374 # CHECK-NEXT:  2,              148  (71.2%)
 375 # CHECK-NEXT:  4,              26  (12.5%)
 376
 377 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 378 # CHECK-NEXT: [# issued], [# cycles]
 379 # CHECK-NEXT:  0,          3  (1.4%)
 380 # CHECK-NEXT:  1,          10  (4.8%)
 381 # CHECK-NEXT:  2,          195  (93.8%)
 382
 383 # CHECK:      Scheduler's queue usage:
 384 # CHECK-NEXT: [1] Resource name.
 385 # CHECK-NEXT: [2] Average number of used buffer entries.
 386 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 387 # CHECK-NEXT: [4] Total number of buffer entries.
 388
 389 # CHECK:       [1]            [2]        [3]        [4]
 390 # CHECK-NEXT: SBPortAny        45         54         54
 391
 392 # CHECK:      Resources:
 393 # CHECK-NEXT: [0]   - SBDivider
 394 # CHECK-NEXT: [1]   - SBFPDivider
 395 # CHECK-NEXT: [2]   - SBPort0
 396 # CHECK-NEXT: [3]   - SBPort1
 397 # CHECK-NEXT: [4]   - SBPort4
 398 # CHECK-NEXT: [5]   - SBPort5
 399 # CHECK-NEXT: [6.0] - SBPort23
 400 # CHECK-NEXT: [6.1] - SBPort23
 401
 402 # CHECK:      Resource pressure per iteration:
 403 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 404 # CHECK-NEXT:  -      -      -      -     2.00    -     2.00   2.00
 405
 406 # CHECK:      Resource pressure by instruction:
 407 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 408 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movq      %rsp, (%rax)
 409 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movq      (%rcx), %rbp
 410 # CHECK-NEXT:  -      -      -      -      -      -     0.95   0.05   movq      (%rdx), %rsi
 411 # CHECK-NEXT:  -      -      -      -     1.00    -     0.05   0.95   movq      %rdi, (%rbx)
 412
 413 # CHECK:      Timeline view:
 414 # CHECK-NEXT: Index     0123456789
 415
 416 # CHECK:      [0,0]     DeER .   .   movq       %rsp, (%rax)
 417 # CHECK-NEXT: [0,1]     DeeeeeER .   movq       (%rcx), %rbp
 418 # CHECK-NEXT: [0,2]     D=eeeeeER.   movq       (%rdx), %rsi
 419 # CHECK-NEXT: [0,3]     D======eER   movq       %rdi, (%rbx)
 420
 421 # CHECK:      Average Wait times (based on the timeline view):
 422 # CHECK-NEXT: [0]: Executions
 423 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 424 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 425 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 426
 427 # CHECK:            [0]    [1]    [2]    [3]
 428 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movq %rsp, (%rax)
 429 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movq (%rcx), %rbp
 430 # CHECK-NEXT: 2.     1     2.0    2.0    0.0       movq (%rdx), %rsi
 431 # CHECK-NEXT: 3.     1     7.0    0.0    0.0       movq %rdi, (%rbx)
 432 # CHECK-NEXT:        1     2.8    1.0    0.0       <total>
 433
 434 # CHECK:      [4] Code Region
 435
 436 # CHECK:      Iterations:        100
 437 # CHECK-NEXT: Instructions:      400
 438 # CHECK-NEXT: Total Cycles:      208
 439 # CHECK-NEXT: Total uOps:        400
 440
 441 # CHECK:      Dispatch Width:    4
 442 # CHECK-NEXT: uOps Per Cycle:    1.92
 443 # CHECK-NEXT: IPC:               1.92
 444 # CHECK-NEXT: Block RThroughput: 2.0
 445
 446 # CHECK:      Instruction Info:
 447 # CHECK-NEXT: [1]: #uOps
 448 # CHECK-NEXT: [2]: Latency
 449 # CHECK-NEXT: [3]: RThroughput
 450 # CHECK-NEXT: [4]: MayLoad
 451 # CHECK-NEXT: [5]: MayStore
 452 # CHECK-NEXT: [6]: HasSideEffects (U)
 453
 454 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 455 # CHECK-NEXT:  1      1     1.00           *      U     movd    %mm0, (%rax)
 456 # CHECK-NEXT:  1      5     0.50    *                   movd    (%rcx), %mm1
 457 # CHECK-NEXT:  1      5     0.50    *                   movd    (%rdx), %mm2
 458 # CHECK-NEXT:  1      1     1.00           *      U     movd    %mm3, (%rbx)
 459
 460 # CHECK:      Dynamic Dispatch Stall Cycles:
 461 # CHECK-NEXT: RAT     - Register unavailable:                      0
 462 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 463 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            147  (70.7%)
 464 # CHECK-NEXT: LQ      - Load queue full:                           0
 465 # CHECK-NEXT: SQ      - Store queue full:                          0
 466 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 467
 468 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 469 # CHECK-NEXT: [# dispatched], [# cycles]
 470 # CHECK-NEXT:  0,              34  (16.3%)
 471 # CHECK-NEXT:  2,              148  (71.2%)
 472 # CHECK-NEXT:  4,              26  (12.5%)
 473
 474 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 475 # CHECK-NEXT: [# issued], [# cycles]
 476 # CHECK-NEXT:  0,          3  (1.4%)
 477 # CHECK-NEXT:  1,          10  (4.8%)
 478 # CHECK-NEXT:  2,          195  (93.8%)
 479
 480 # CHECK:      Scheduler's queue usage:
 481 # CHECK-NEXT: [1] Resource name.
 482 # CHECK-NEXT: [2] Average number of used buffer entries.
 483 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 484 # CHECK-NEXT: [4] Total number of buffer entries.
 485
 486 # CHECK:       [1]            [2]        [3]        [4]
 487 # CHECK-NEXT: SBPortAny        45         54         54
 488
 489 # CHECK:      Resources:
 490 # CHECK-NEXT: [0]   - SBDivider
 491 # CHECK-NEXT: [1]   - SBFPDivider
 492 # CHECK-NEXT: [2]   - SBPort0
 493 # CHECK-NEXT: [3]   - SBPort1
 494 # CHECK-NEXT: [4]   - SBPort4
 495 # CHECK-NEXT: [5]   - SBPort5
 496 # CHECK-NEXT: [6.0] - SBPort23
 497 # CHECK-NEXT: [6.1] - SBPort23
 498
 499 # CHECK:      Resource pressure per iteration:
 500 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 501 # CHECK-NEXT:  -      -      -      -     2.00    -     2.00   2.00
 502
 503 # CHECK:      Resource pressure by instruction:
 504 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 505 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movd      %mm0, (%rax)
 506 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movd      (%rcx), %mm1
 507 # CHECK-NEXT:  -      -      -      -      -      -     0.95   0.05   movd      (%rdx), %mm2
 508 # CHECK-NEXT:  -      -      -      -     1.00    -     0.05   0.95   movd      %mm3, (%rbx)
 509
 510 # CHECK:      Timeline view:
 511 # CHECK-NEXT: Index     0123456789
 512
 513 # CHECK:      [0,0]     DeER .   .   movd       %mm0, (%rax)
 514 # CHECK-NEXT: [0,1]     DeeeeeER .   movd       (%rcx), %mm1
 515 # CHECK-NEXT: [0,2]     D=eeeeeER.   movd       (%rdx), %mm2
 516 # CHECK-NEXT: [0,3]     D======eER   movd       %mm3, (%rbx)
 517
 518 # CHECK:      Average Wait times (based on the timeline view):
 519 # CHECK-NEXT: [0]: Executions
 520 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 521 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 522 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 523
 524 # CHECK:            [0]    [1]    [2]    [3]
 525 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movd %mm0, (%rax)
 526 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movd (%rcx), %mm1
 527 # CHECK-NEXT: 2.     1     2.0    2.0    0.0       movd (%rdx), %mm2
 528 # CHECK-NEXT: 3.     1     7.0    0.0    0.0       movd %mm3, (%rbx)
 529 # CHECK-NEXT:        1     2.8    1.0    0.0       <total>
 530
 531 # CHECK:      [5] Code Region
 532
 533 # CHECK:      Iterations:        100
 534 # CHECK-NEXT: Instructions:      400
 535 # CHECK-NEXT: Total Cycles:      209
 536 # CHECK-NEXT: Total uOps:        400
 537
 538 # CHECK:      Dispatch Width:    4
 539 # CHECK-NEXT: uOps Per Cycle:    1.91
 540 # CHECK-NEXT: IPC:               1.91
 541 # CHECK-NEXT: Block RThroughput: 2.0
 542
 543 # CHECK:      Instruction Info:
 544 # CHECK-NEXT: [1]: #uOps
 545 # CHECK-NEXT: [2]: Latency
 546 # CHECK-NEXT: [3]: RThroughput
 547 # CHECK-NEXT: [4]: MayLoad
 548 # CHECK-NEXT: [5]: MayStore
 549 # CHECK-NEXT: [6]: HasSideEffects (U)
 550
 551 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 552 # CHECK-NEXT:  1      1     1.00           *            movaps  %xmm0, (%rax)
 553 # CHECK-NEXT:  1      6     0.50    *                   movaps  (%rcx), %xmm1
 554 # CHECK-NEXT:  1      6     0.50    *                   movaps  (%rdx), %xmm2
 555 # CHECK-NEXT:  1      1     1.00           *            movaps  %xmm3, (%rbx)
 556
 557 # CHECK:      Dynamic Dispatch Stall Cycles:
 558 # CHECK-NEXT: RAT     - Register unavailable:                      0
 559 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 560 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            147  (70.3%)
 561 # CHECK-NEXT: LQ      - Load queue full:                           0
 562 # CHECK-NEXT: SQ      - Store queue full:                          0
 563 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 564
 565 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 566 # CHECK-NEXT: [# dispatched], [# cycles]
 567 # CHECK-NEXT:  0,              35  (16.7%)
 568 # CHECK-NEXT:  2,              148  (70.8%)
 569 # CHECK-NEXT:  4,              26  (12.4%)
 570
 571 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 572 # CHECK-NEXT: [# issued], [# cycles]
 573 # CHECK-NEXT:  0,          3  (1.4%)
 574 # CHECK-NEXT:  1,          12  (5.7%)
 575 # CHECK-NEXT:  2,          194  (92.8%)
 576
 577 # CHECK:      Scheduler's queue usage:
 578 # CHECK-NEXT: [1] Resource name.
 579 # CHECK-NEXT: [2] Average number of used buffer entries.
 580 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 581 # CHECK-NEXT: [4] Total number of buffer entries.
 582
 583 # CHECK:       [1]            [2]        [3]        [4]
 584 # CHECK-NEXT: SBPortAny        45         54         54
 585
 586 # CHECK:      Resources:
 587 # CHECK-NEXT: [0]   - SBDivider
 588 # CHECK-NEXT: [1]   - SBFPDivider
 589 # CHECK-NEXT: [2]   - SBPort0
 590 # CHECK-NEXT: [3]   - SBPort1
 591 # CHECK-NEXT: [4]   - SBPort4
 592 # CHECK-NEXT: [5]   - SBPort5
 593 # CHECK-NEXT: [6.0] - SBPort23
 594 # CHECK-NEXT: [6.1] - SBPort23
 595
 596 # CHECK:      Resource pressure per iteration:
 597 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]
 598 # CHECK-NEXT:  -      -      -      -     2.00    -     2.00   2.00
 599
 600 # CHECK:      Resource pressure by instruction:
 601 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6.0]  [6.1]  Instructions:
 602 # CHECK-NEXT:  -      -      -      -     1.00    -      -     1.00   movaps    %xmm0, (%rax)
 603 # CHECK-NEXT:  -      -      -      -      -      -     1.00    -     movaps    (%rcx), %xmm1
 604 # CHECK-NEXT:  -      -      -      -      -      -     0.94   0.06   movaps    (%rdx), %xmm2
 605 # CHECK-NEXT:  -      -      -      -     1.00    -     0.06   0.94   movaps    %xmm3, (%rbx)
 606
 607 # CHECK:      Timeline view:
 608 # CHECK-NEXT:                     0
 609 # CHECK-NEXT: Index     0123456789
 610
 611 # CHECK:      [0,0]     DeER .    .   movaps    %xmm0, (%rax)
 612 # CHECK-NEXT: [0,1]     DeeeeeeER .   movaps    (%rcx), %xmm1
 613 # CHECK-NEXT: [0,2]     D=eeeeeeER.   movaps    (%rdx), %xmm2
 614 # CHECK-NEXT: [0,3]     D=======eER   movaps    %xmm3, (%rbx)
 615
 616 # CHECK:      Average Wait times (based on the timeline view):
 617 # CHECK-NEXT: [0]: Executions
 618 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 619 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 620 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 621
 622 # CHECK:            [0]    [1]    [2]    [3]
 623 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movaps       %xmm0, (%rax)
 624 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movaps       (%rcx), %xmm1
 625 # CHECK-NEXT: 2.     1     2.0    2.0    0.0       movaps       (%rdx), %xmm2
 626 # CHECK-NEXT: 3.     1     8.0    0.0    0.0       movaps       %xmm3, (%rbx)
 627 # CHECK-NEXT:        1     3.0    1.0    0.0       <total>