llvm/test/tools/llvm-mca/X86/BdVer2/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=bdver2 -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 # LLVM-MCA-BEGIN
  47 vmovaps (%rax), %ymm0
  48 vmovaps (%rcx), %ymm1
  49 vmovaps (%rdx), %ymm2
  50 vmovaps (%rbx), %ymm3
  51 # LLVM-MCA-END
  52
  53 # CHECK:      [0] Code Region
  54
  55 # CHECK:      Iterations:        100
  56 # CHECK-NEXT: Instructions:      400
  57 # CHECK-NEXT: Total Cycles:      406
  58 # CHECK-NEXT: Total uOps:        400
  59
  60 # CHECK:      Dispatch Width:    4
  61 # CHECK-NEXT: uOps Per Cycle:    0.99
  62 # CHECK-NEXT: IPC:               0.99
  63 # CHECK-NEXT: Block RThroughput: 4.0
  64
  65 # CHECK:      Instruction Info:
  66 # CHECK-NEXT: [1]: #uOps
  67 # CHECK-NEXT: [2]: Latency
  68 # CHECK-NEXT: [3]: RThroughput
  69 # CHECK-NEXT: [4]: MayLoad
  70 # CHECK-NEXT: [5]: MayStore
  71 # CHECK-NEXT: [6]: HasSideEffects (U)
  72
  73 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
  74 # CHECK-NEXT:  1      5     1.00    *                   movb    (%rax), %spl
  75 # CHECK-NEXT:  1      5     1.00    *                   movb    (%rcx), %bpl
  76 # CHECK-NEXT:  1      5     1.00    *                   movb    (%rdx), %sil
  77 # CHECK-NEXT:  1      5     1.00    *                   movb    (%rbx), %dil
  78
  79 # CHECK:      Dynamic Dispatch Stall Cycles:
  80 # CHECK-NEXT: RAT     - Register unavailable:                      0
  81 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
  82 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            0
  83 # CHECK-NEXT: LQ      - Load queue full:                           354  (87.2%)
  84 # CHECK-NEXT: SQ      - Store queue full:                          0
  85 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
  86 # CHECK-NEXT: USH     - Uncategorised Structural Hazard:           0
  87
  88 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
  89 # CHECK-NEXT: [# dispatched], [# cycles]
  90 # CHECK-NEXT:  0,              217  (53.4%)
  91 # CHECK-NEXT:  2,              178  (43.8%)
  92 # CHECK-NEXT:  4,              11  (2.7%)
  93
  94 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
  95 # CHECK-NEXT: [# issued], [# cycles]
  96 # CHECK-NEXT:  0,          206  (50.7%)
  97 # CHECK-NEXT:  2,          200  (49.3%)
  98
  99 # CHECK:      Scheduler's queue usage:
 100 # CHECK-NEXT: [1] Resource name.
 101 # CHECK-NEXT: [2] Average number of used buffer entries.
 102 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 103 # CHECK-NEXT: [4] Total number of buffer entries.
 104
 105 # CHECK:       [1]            [2]        [3]        [4]
 106 # CHECK-NEXT: PdEX             31         34         40
 107 # CHECK-NEXT: PdFPU            0          0          64
 108 # CHECK-NEXT: PdLoad           36         40         40
 109 # CHECK-NEXT: PdStore          0          0          24
 110
 111 # CHECK:      Resources:
 112 # CHECK-NEXT: [0.0] - PdAGLU01
 113 # CHECK-NEXT: [0.1] - PdAGLU01
 114 # CHECK-NEXT: [1]   - PdBranch
 115 # CHECK-NEXT: [2]   - PdCount
 116 # CHECK-NEXT: [3]   - PdDiv
 117 # CHECK-NEXT: [4]   - PdEX0
 118 # CHECK-NEXT: [5]   - PdEX1
 119 # CHECK-NEXT: [6]   - PdFPCVT
 120 # CHECK-NEXT: [7.0] - PdFPFMA
 121 # CHECK-NEXT: [7.1] - PdFPFMA
 122 # CHECK-NEXT: [8.0] - PdFPMAL
 123 # CHECK-NEXT: [8.1] - PdFPMAL
 124 # CHECK-NEXT: [9]   - PdFPMMA
 125 # CHECK-NEXT: [10]  - PdFPSTO
 126 # CHECK-NEXT: [11]  - PdFPU0
 127 # CHECK-NEXT: [12]  - PdFPU1
 128 # CHECK-NEXT: [13]  - PdFPU2
 129 # CHECK-NEXT: [14]  - PdFPU3
 130 # CHECK-NEXT: [15]  - PdFPXBR
 131 # CHECK-NEXT: [16.0] - PdLoad
 132 # CHECK-NEXT: [16.1] - PdLoad
 133 # CHECK-NEXT: [17]  - PdMul
 134 # CHECK-NEXT: [18]  - PdStore
 135
 136 # CHECK:      Resource pressure per iteration:
 137 # CHECK-NEXT: [0.0]  [0.1]  [1]    [2]    [3]    [4]    [5]    [6]    [7.0]  [7.1]  [8.0]  [8.1]  [9]    [10]   [11]   [12]   [13]   [14]   [15]   [16.0] [16.1] [17]   [18]
 138 # CHECK-NEXT: 4.00   4.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     4.00   4.00    -      -
 139
 140 # CHECK:      Resource pressure by instruction:
 141 # CHECK-NEXT: [0.0]  [0.1]  [1]    [2]    [3]    [4]    [5]    [6]    [7.0]  [7.1]  [8.0]  [8.1]  [9]    [10]   [11]   [12]   [13]   [14]   [15]   [16.0] [16.1] [17]   [18]   Instructions:
 142 # CHECK-NEXT:  -     2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -     movb     (%rax), %spl
 143 # CHECK-NEXT: 2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -      -     movb     (%rcx), %bpl
 144 # CHECK-NEXT:  -     2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -     movb     (%rdx), %sil
 145 # CHECK-NEXT: 2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -      -     movb     (%rbx), %dil
 146
 147 # CHECK:      Timeline view:
 148 # CHECK-NEXT: Index     0123456789
 149
 150 # CHECK:      [0,0]     DeeeeeER .   movb       (%rax), %spl
 151 # CHECK-NEXT: [0,1]     DeeeeeER .   movb       (%rcx), %bpl
 152 # CHECK-NEXT: [0,2]     D==eeeeeER   movb       (%rdx), %sil
 153 # CHECK-NEXT: [0,3]     D==eeeeeER   movb       (%rbx), %dil
 154
 155 # CHECK:      Average Wait times (based on the timeline view):
 156 # CHECK-NEXT: [0]: Executions
 157 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 158 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 159 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 160
 161 # CHECK:            [0]    [1]    [2]    [3]
 162 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movb (%rax), %spl
 163 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movb (%rcx), %bpl
 164 # CHECK-NEXT: 2.     1     3.0    3.0    0.0       movb (%rdx), %sil
 165 # CHECK-NEXT: 3.     1     3.0    3.0    0.0       movb (%rbx), %dil
 166 # CHECK-NEXT:        1     2.0    2.0    0.0       <total>
 167
 168 # CHECK:      [1] Code Region
 169
 170 # CHECK:      Iterations:        100
 171 # CHECK-NEXT: Instructions:      400
 172 # CHECK-NEXT: Total Cycles:      406
 173 # CHECK-NEXT: Total uOps:        400
 174
 175 # CHECK:      Dispatch Width:    4
 176 # CHECK-NEXT: uOps Per Cycle:    0.99
 177 # CHECK-NEXT: IPC:               0.99
 178 # CHECK-NEXT: Block RThroughput: 4.0
 179
 180 # CHECK:      Instruction Info:
 181 # CHECK-NEXT: [1]: #uOps
 182 # CHECK-NEXT: [2]: Latency
 183 # CHECK-NEXT: [3]: RThroughput
 184 # CHECK-NEXT: [4]: MayLoad
 185 # CHECK-NEXT: [5]: MayStore
 186 # CHECK-NEXT: [6]: HasSideEffects (U)
 187
 188 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 189 # CHECK-NEXT:  1      5     1.00    *                   movw    (%rax), %sp
 190 # CHECK-NEXT:  1      5     1.00    *                   movw    (%rcx), %bp
 191 # CHECK-NEXT:  1      5     1.00    *                   movw    (%rdx), %si
 192 # CHECK-NEXT:  1      5     1.00    *                   movw    (%rbx), %di
 193
 194 # CHECK:      Dynamic Dispatch Stall Cycles:
 195 # CHECK-NEXT: RAT     - Register unavailable:                      0
 196 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 197 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            0
 198 # CHECK-NEXT: LQ      - Load queue full:                           354  (87.2%)
 199 # CHECK-NEXT: SQ      - Store queue full:                          0
 200 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 201 # CHECK-NEXT: USH     - Uncategorised Structural Hazard:           0
 202
 203 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 204 # CHECK-NEXT: [# dispatched], [# cycles]
 205 # CHECK-NEXT:  0,              217  (53.4%)
 206 # CHECK-NEXT:  2,              178  (43.8%)
 207 # CHECK-NEXT:  4,              11  (2.7%)
 208
 209 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 210 # CHECK-NEXT: [# issued], [# cycles]
 211 # CHECK-NEXT:  0,          206  (50.7%)
 212 # CHECK-NEXT:  2,          200  (49.3%)
 213
 214 # CHECK:      Scheduler's queue usage:
 215 # CHECK-NEXT: [1] Resource name.
 216 # CHECK-NEXT: [2] Average number of used buffer entries.
 217 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 218 # CHECK-NEXT: [4] Total number of buffer entries.
 219
 220 # CHECK:       [1]            [2]        [3]        [4]
 221 # CHECK-NEXT: PdEX             31         34         40
 222 # CHECK-NEXT: PdFPU            0          0          64
 223 # CHECK-NEXT: PdLoad           36         40         40
 224 # CHECK-NEXT: PdStore          0          0          24
 225
 226 # CHECK:      Resources:
 227 # CHECK-NEXT: [0.0] - PdAGLU01
 228 # CHECK-NEXT: [0.1] - PdAGLU01
 229 # CHECK-NEXT: [1]   - PdBranch
 230 # CHECK-NEXT: [2]   - PdCount
 231 # CHECK-NEXT: [3]   - PdDiv
 232 # CHECK-NEXT: [4]   - PdEX0
 233 # CHECK-NEXT: [5]   - PdEX1
 234 # CHECK-NEXT: [6]   - PdFPCVT
 235 # CHECK-NEXT: [7.0] - PdFPFMA
 236 # CHECK-NEXT: [7.1] - PdFPFMA
 237 # CHECK-NEXT: [8.0] - PdFPMAL
 238 # CHECK-NEXT: [8.1] - PdFPMAL
 239 # CHECK-NEXT: [9]   - PdFPMMA
 240 # CHECK-NEXT: [10]  - PdFPSTO
 241 # CHECK-NEXT: [11]  - PdFPU0
 242 # CHECK-NEXT: [12]  - PdFPU1
 243 # CHECK-NEXT: [13]  - PdFPU2
 244 # CHECK-NEXT: [14]  - PdFPU3
 245 # CHECK-NEXT: [15]  - PdFPXBR
 246 # CHECK-NEXT: [16.0] - PdLoad
 247 # CHECK-NEXT: [16.1] - PdLoad
 248 # CHECK-NEXT: [17]  - PdMul
 249 # CHECK-NEXT: [18]  - PdStore
 250
 251 # CHECK:      Resource pressure per iteration:
 252 # CHECK-NEXT: [0.0]  [0.1]  [1]    [2]    [3]    [4]    [5]    [6]    [7.0]  [7.1]  [8.0]  [8.1]  [9]    [10]   [11]   [12]   [13]   [14]   [15]   [16.0] [16.1] [17]   [18]
 253 # CHECK-NEXT: 4.00   4.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     4.00   4.00    -      -
 254
 255 # CHECK:      Resource pressure by instruction:
 256 # CHECK-NEXT: [0.0]  [0.1]  [1]    [2]    [3]    [4]    [5]    [6]    [7.0]  [7.1]  [8.0]  [8.1]  [9]    [10]   [11]   [12]   [13]   [14]   [15]   [16.0] [16.1] [17]   [18]   Instructions:
 257 # CHECK-NEXT:  -     2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -     movw     (%rax), %sp
 258 # CHECK-NEXT: 2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -      -     movw     (%rcx), %bp
 259 # CHECK-NEXT:  -     2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -     movw     (%rdx), %si
 260 # CHECK-NEXT: 2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -      -     movw     (%rbx), %di
 261
 262 # CHECK:      Timeline view:
 263 # CHECK-NEXT: Index     0123456789
 264
 265 # CHECK:      [0,0]     DeeeeeER .   movw       (%rax), %sp
 266 # CHECK-NEXT: [0,1]     DeeeeeER .   movw       (%rcx), %bp
 267 # CHECK-NEXT: [0,2]     D==eeeeeER   movw       (%rdx), %si
 268 # CHECK-NEXT: [0,3]     D==eeeeeER   movw       (%rbx), %di
 269
 270 # CHECK:      Average Wait times (based on the timeline view):
 271 # CHECK-NEXT: [0]: Executions
 272 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 273 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 274 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 275
 276 # CHECK:            [0]    [1]    [2]    [3]
 277 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movw (%rax), %sp
 278 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movw (%rcx), %bp
 279 # CHECK-NEXT: 2.     1     3.0    3.0    0.0       movw (%rdx), %si
 280 # CHECK-NEXT: 3.     1     3.0    3.0    0.0       movw (%rbx), %di
 281 # CHECK-NEXT:        1     2.0    2.0    0.0       <total>
 282
 283 # CHECK:      [2] Code Region
 284
 285 # CHECK:      Iterations:        100
 286 # CHECK-NEXT: Instructions:      400
 287 # CHECK-NEXT: Total Cycles:      406
 288 # CHECK-NEXT: Total uOps:        400
 289
 290 # CHECK:      Dispatch Width:    4
 291 # CHECK-NEXT: uOps Per Cycle:    0.99
 292 # CHECK-NEXT: IPC:               0.99
 293 # CHECK-NEXT: Block RThroughput: 4.0
 294
 295 # CHECK:      Instruction Info:
 296 # CHECK-NEXT: [1]: #uOps
 297 # CHECK-NEXT: [2]: Latency
 298 # CHECK-NEXT: [3]: RThroughput
 299 # CHECK-NEXT: [4]: MayLoad
 300 # CHECK-NEXT: [5]: MayStore
 301 # CHECK-NEXT: [6]: HasSideEffects (U)
 302
 303 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 304 # CHECK-NEXT:  1      5     1.00    *                   movl    (%rax), %esp
 305 # CHECK-NEXT:  1      5     1.00    *                   movl    (%rcx), %ebp
 306 # CHECK-NEXT:  1      5     1.00    *                   movl    (%rdx), %esi
 307 # CHECK-NEXT:  1      5     1.00    *                   movl    (%rbx), %edi
 308
 309 # CHECK:      Dynamic Dispatch Stall Cycles:
 310 # CHECK-NEXT: RAT     - Register unavailable:                      0
 311 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 312 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            0
 313 # CHECK-NEXT: LQ      - Load queue full:                           354  (87.2%)
 314 # CHECK-NEXT: SQ      - Store queue full:                          0
 315 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 316 # CHECK-NEXT: USH     - Uncategorised Structural Hazard:           0
 317
 318 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 319 # CHECK-NEXT: [# dispatched], [# cycles]
 320 # CHECK-NEXT:  0,              217  (53.4%)
 321 # CHECK-NEXT:  2,              178  (43.8%)
 322 # CHECK-NEXT:  4,              11  (2.7%)
 323
 324 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 325 # CHECK-NEXT: [# issued], [# cycles]
 326 # CHECK-NEXT:  0,          206  (50.7%)
 327 # CHECK-NEXT:  2,          200  (49.3%)
 328
 329 # CHECK:      Scheduler's queue usage:
 330 # CHECK-NEXT: [1] Resource name.
 331 # CHECK-NEXT: [2] Average number of used buffer entries.
 332 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 333 # CHECK-NEXT: [4] Total number of buffer entries.
 334
 335 # CHECK:       [1]            [2]        [3]        [4]
 336 # CHECK-NEXT: PdEX             31         34         40
 337 # CHECK-NEXT: PdFPU            0          0          64
 338 # CHECK-NEXT: PdLoad           36         40         40
 339 # CHECK-NEXT: PdStore          0          0          24
 340
 341 # CHECK:      Resources:
 342 # CHECK-NEXT: [0.0] - PdAGLU01
 343 # CHECK-NEXT: [0.1] - PdAGLU01
 344 # CHECK-NEXT: [1]   - PdBranch
 345 # CHECK-NEXT: [2]   - PdCount
 346 # CHECK-NEXT: [3]   - PdDiv
 347 # CHECK-NEXT: [4]   - PdEX0
 348 # CHECK-NEXT: [5]   - PdEX1
 349 # CHECK-NEXT: [6]   - PdFPCVT
 350 # CHECK-NEXT: [7.0] - PdFPFMA
 351 # CHECK-NEXT: [7.1] - PdFPFMA
 352 # CHECK-NEXT: [8.0] - PdFPMAL
 353 # CHECK-NEXT: [8.1] - PdFPMAL
 354 # CHECK-NEXT: [9]   - PdFPMMA
 355 # CHECK-NEXT: [10]  - PdFPSTO
 356 # CHECK-NEXT: [11]  - PdFPU0
 357 # CHECK-NEXT: [12]  - PdFPU1
 358 # CHECK-NEXT: [13]  - PdFPU2
 359 # CHECK-NEXT: [14]  - PdFPU3
 360 # CHECK-NEXT: [15]  - PdFPXBR
 361 # CHECK-NEXT: [16.0] - PdLoad
 362 # CHECK-NEXT: [16.1] - PdLoad
 363 # CHECK-NEXT: [17]  - PdMul
 364 # CHECK-NEXT: [18]  - PdStore
 365
 366 # CHECK:      Resource pressure per iteration:
 367 # CHECK-NEXT: [0.0]  [0.1]  [1]    [2]    [3]    [4]    [5]    [6]    [7.0]  [7.1]  [8.0]  [8.1]  [9]    [10]   [11]   [12]   [13]   [14]   [15]   [16.0] [16.1] [17]   [18]
 368 # CHECK-NEXT: 4.00   4.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     4.00   4.00    -      -
 369
 370 # CHECK:      Resource pressure by instruction:
 371 # CHECK-NEXT: [0.0]  [0.1]  [1]    [2]    [3]    [4]    [5]    [6]    [7.0]  [7.1]  [8.0]  [8.1]  [9]    [10]   [11]   [12]   [13]   [14]   [15]   [16.0] [16.1] [17]   [18]   Instructions:
 372 # CHECK-NEXT:  -     2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -     movl     (%rax), %esp
 373 # CHECK-NEXT: 2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -      -     movl     (%rcx), %ebp
 374 # CHECK-NEXT:  -     2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -     movl     (%rdx), %esi
 375 # CHECK-NEXT: 2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -      -     movl     (%rbx), %edi
 376
 377 # CHECK:      Timeline view:
 378 # CHECK-NEXT: Index     0123456789
 379
 380 # CHECK:      [0,0]     DeeeeeER .   movl       (%rax), %esp
 381 # CHECK-NEXT: [0,1]     DeeeeeER .   movl       (%rcx), %ebp
 382 # CHECK-NEXT: [0,2]     D==eeeeeER   movl       (%rdx), %esi
 383 # CHECK-NEXT: [0,3]     D==eeeeeER   movl       (%rbx), %edi
 384
 385 # CHECK:      Average Wait times (based on the timeline view):
 386 # CHECK-NEXT: [0]: Executions
 387 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 388 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 389 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 390
 391 # CHECK:            [0]    [1]    [2]    [3]
 392 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movl (%rax), %esp
 393 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movl (%rcx), %ebp
 394 # CHECK-NEXT: 2.     1     3.0    3.0    0.0       movl (%rdx), %esi
 395 # CHECK-NEXT: 3.     1     3.0    3.0    0.0       movl (%rbx), %edi
 396 # CHECK-NEXT:        1     2.0    2.0    0.0       <total>
 397
 398 # CHECK:      [3] Code Region
 399
 400 # CHECK:      Iterations:        100
 401 # CHECK-NEXT: Instructions:      400
 402 # CHECK-NEXT: Total Cycles:      406
 403 # CHECK-NEXT: Total uOps:        400
 404
 405 # CHECK:      Dispatch Width:    4
 406 # CHECK-NEXT: uOps Per Cycle:    0.99
 407 # CHECK-NEXT: IPC:               0.99
 408 # CHECK-NEXT: Block RThroughput: 4.0
 409
 410 # CHECK:      Instruction Info:
 411 # CHECK-NEXT: [1]: #uOps
 412 # CHECK-NEXT: [2]: Latency
 413 # CHECK-NEXT: [3]: RThroughput
 414 # CHECK-NEXT: [4]: MayLoad
 415 # CHECK-NEXT: [5]: MayStore
 416 # CHECK-NEXT: [6]: HasSideEffects (U)
 417
 418 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 419 # CHECK-NEXT:  1      5     1.00    *                   movq    (%rax), %rsp
 420 # CHECK-NEXT:  1      5     1.00    *                   movq    (%rcx), %rbp
 421 # CHECK-NEXT:  1      5     1.00    *                   movq    (%rdx), %rsi
 422 # CHECK-NEXT:  1      5     1.00    *                   movq    (%rbx), %rdi
 423
 424 # CHECK:      Dynamic Dispatch Stall Cycles:
 425 # CHECK-NEXT: RAT     - Register unavailable:                      0
 426 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 427 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            0
 428 # CHECK-NEXT: LQ      - Load queue full:                           354  (87.2%)
 429 # CHECK-NEXT: SQ      - Store queue full:                          0
 430 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 431 # CHECK-NEXT: USH     - Uncategorised Structural Hazard:           0
 432
 433 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 434 # CHECK-NEXT: [# dispatched], [# cycles]
 435 # CHECK-NEXT:  0,              217  (53.4%)
 436 # CHECK-NEXT:  2,              178  (43.8%)
 437 # CHECK-NEXT:  4,              11  (2.7%)
 438
 439 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 440 # CHECK-NEXT: [# issued], [# cycles]
 441 # CHECK-NEXT:  0,          206  (50.7%)
 442 # CHECK-NEXT:  2,          200  (49.3%)
 443
 444 # CHECK:      Scheduler's queue usage:
 445 # CHECK-NEXT: [1] Resource name.
 446 # CHECK-NEXT: [2] Average number of used buffer entries.
 447 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 448 # CHECK-NEXT: [4] Total number of buffer entries.
 449
 450 # CHECK:       [1]            [2]        [3]        [4]
 451 # CHECK-NEXT: PdEX             31         34         40
 452 # CHECK-NEXT: PdFPU            0          0          64
 453 # CHECK-NEXT: PdLoad           36         40         40
 454 # CHECK-NEXT: PdStore          0          0          24
 455
 456 # CHECK:      Resources:
 457 # CHECK-NEXT: [0.0] - PdAGLU01
 458 # CHECK-NEXT: [0.1] - PdAGLU01
 459 # CHECK-NEXT: [1]   - PdBranch
 460 # CHECK-NEXT: [2]   - PdCount
 461 # CHECK-NEXT: [3]   - PdDiv
 462 # CHECK-NEXT: [4]   - PdEX0
 463 # CHECK-NEXT: [5]   - PdEX1
 464 # CHECK-NEXT: [6]   - PdFPCVT
 465 # CHECK-NEXT: [7.0] - PdFPFMA
 466 # CHECK-NEXT: [7.1] - PdFPFMA
 467 # CHECK-NEXT: [8.0] - PdFPMAL
 468 # CHECK-NEXT: [8.1] - PdFPMAL
 469 # CHECK-NEXT: [9]   - PdFPMMA
 470 # CHECK-NEXT: [10]  - PdFPSTO
 471 # CHECK-NEXT: [11]  - PdFPU0
 472 # CHECK-NEXT: [12]  - PdFPU1
 473 # CHECK-NEXT: [13]  - PdFPU2
 474 # CHECK-NEXT: [14]  - PdFPU3
 475 # CHECK-NEXT: [15]  - PdFPXBR
 476 # CHECK-NEXT: [16.0] - PdLoad
 477 # CHECK-NEXT: [16.1] - PdLoad
 478 # CHECK-NEXT: [17]  - PdMul
 479 # CHECK-NEXT: [18]  - PdStore
 480
 481 # CHECK:      Resource pressure per iteration:
 482 # CHECK-NEXT: [0.0]  [0.1]  [1]    [2]    [3]    [4]    [5]    [6]    [7.0]  [7.1]  [8.0]  [8.1]  [9]    [10]   [11]   [12]   [13]   [14]   [15]   [16.0] [16.1] [17]   [18]
 483 # CHECK-NEXT: 4.00   4.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     4.00   4.00    -      -
 484
 485 # CHECK:      Resource pressure by instruction:
 486 # CHECK-NEXT: [0.0]  [0.1]  [1]    [2]    [3]    [4]    [5]    [6]    [7.0]  [7.1]  [8.0]  [8.1]  [9]    [10]   [11]   [12]   [13]   [14]   [15]   [16.0] [16.1] [17]   [18]   Instructions:
 487 # CHECK-NEXT:  -     2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -     movq     (%rax), %rsp
 488 # CHECK-NEXT: 2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -      -     movq     (%rcx), %rbp
 489 # CHECK-NEXT:  -     2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -     movq     (%rdx), %rsi
 490 # CHECK-NEXT: 2.00    -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -      -     2.00    -      -      -     movq     (%rbx), %rdi
 491
 492 # CHECK:      Timeline view:
 493 # CHECK-NEXT: Index     0123456789
 494
 495 # CHECK:      [0,0]     DeeeeeER .   movq       (%rax), %rsp
 496 # CHECK-NEXT: [0,1]     DeeeeeER .   movq       (%rcx), %rbp
 497 # CHECK-NEXT: [0,2]     D==eeeeeER   movq       (%rdx), %rsi
 498 # CHECK-NEXT: [0,3]     D==eeeeeER   movq       (%rbx), %rdi
 499
 500 # CHECK:      Average Wait times (based on the timeline view):
 501 # CHECK-NEXT: [0]: Executions
 502 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 503 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 504 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 505
 506 # CHECK:            [0]    [1]    [2]    [3]
 507 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movq (%rax), %rsp
 508 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movq (%rcx), %rbp
 509 # CHECK-NEXT: 2.     1     3.0    3.0    0.0       movq (%rdx), %rsi
 510 # CHECK-NEXT: 3.     1     3.0    3.0    0.0       movq (%rbx), %rdi
 511 # CHECK-NEXT:        1     2.0    2.0    0.0       <total>
 512
 513 # CHECK:      [4] Code Region
 514
 515 # CHECK:      Iterations:        100
 516 # CHECK-NEXT: Instructions:      400
 517 # CHECK-NEXT: Total Cycles:      605
 518 # CHECK-NEXT: Total uOps:        400
 519
 520 # CHECK:      Dispatch Width:    4
 521 # CHECK-NEXT: uOps Per Cycle:    0.66
 522 # CHECK-NEXT: IPC:               0.66
 523 # CHECK-NEXT: Block RThroughput: 6.0
 524
 525 # CHECK:      Instruction Info:
 526 # CHECK-NEXT: [1]: #uOps
 527 # CHECK-NEXT: [2]: Latency
 528 # CHECK-NEXT: [3]: RThroughput
 529 # CHECK-NEXT: [4]: MayLoad
 530 # CHECK-NEXT: [5]: MayStore
 531 # CHECK-NEXT: [6]: HasSideEffects (U)
 532
 533 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 534 # CHECK-NEXT:  1      5     1.50    *                   movd    (%rax), %mm0
 535 # CHECK-NEXT:  1      5     1.50    *                   movd    (%rcx), %mm1
 536 # CHECK-NEXT:  1      5     1.50    *                   movd    (%rdx), %mm2
 537 # CHECK-NEXT:  1      5     1.50    *                   movd    (%rbx), %mm3
 538
 539 # CHECK:      Dynamic Dispatch Stall Cycles:
 540 # CHECK-NEXT: RAT     - Register unavailable:                      0
 541 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 542 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            0
 543 # CHECK-NEXT: LQ      - Load queue full:                           533  (88.1%)
 544 # CHECK-NEXT: SQ      - Store queue full:                          0
 545 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 546 # CHECK-NEXT: USH     - Uncategorised Structural Hazard:           0
 547
 548 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 549 # CHECK-NEXT: [# dispatched], [# cycles]
 550 # CHECK-NEXT:  0,              416  (68.8%)
 551 # CHECK-NEXT:  2,              178  (29.4%)
 552 # CHECK-NEXT:  4,              11  (1.8%)
 553
 554 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 555 # CHECK-NEXT: [# issued], [# cycles]
 556 # CHECK-NEXT:  0,          405  (66.9%)
 557 # CHECK-NEXT:  2,          200  (33.1%)
 558
 559 # CHECK:      Scheduler's queue usage:
 560 # CHECK-NEXT: [1] Resource name.
 561 # CHECK-NEXT: [2] Average number of used buffer entries.
 562 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 563 # CHECK-NEXT: [4] Total number of buffer entries.
 564
 565 # CHECK:       [1]            [2]        [3]        [4]
 566 # CHECK-NEXT: PdEX             33         36         40
 567 # CHECK-NEXT: PdFPU            33         36         64
 568 # CHECK-NEXT: PdLoad           37         40         40
 569 # CHECK-NEXT: PdStore          0          0          24
 570
 571 # CHECK:      Resources:
 572 # CHECK-NEXT: [0.0] - PdAGLU01
 573 # CHECK-NEXT: [0.1] - PdAGLU01
 574 # CHECK-NEXT: [1]   - PdBranch
 575 # CHECK-NEXT: [2]   - PdCount
 576 # CHECK-NEXT: [3]   - PdDiv
 577 # CHECK-NEXT: [4]   - PdEX0
 578 # CHECK-NEXT: [5]   - PdEX1
 579 # CHECK-NEXT: [6]   - PdFPCVT
 580 # CHECK-NEXT: [7.0] - PdFPFMA
 581 # CHECK-NEXT: [7.1] - PdFPFMA
 582 # CHECK-NEXT: [8.0] - PdFPMAL
 583 # CHECK-NEXT: [8.1] - PdFPMAL
 584 # CHECK-NEXT: [9]   - PdFPMMA
 585 # CHECK-NEXT: [10]  - PdFPSTO
 586 # CHECK-NEXT: [11]  - PdFPU0
 587 # CHECK-NEXT: [12]  - PdFPU1
 588 # CHECK-NEXT: [13]  - PdFPU2
 589 # CHECK-NEXT: [14]  - PdFPU3
 590 # CHECK-NEXT: [15]  - PdFPXBR
 591 # CHECK-NEXT: [16.0] - PdLoad
 592 # CHECK-NEXT: [16.1] - PdLoad
 593 # CHECK-NEXT: [17]  - PdMul
 594 # CHECK-NEXT: [18]  - PdStore
 595
 596 # CHECK:      Resource pressure per iteration:
 597 # CHECK-NEXT: [0.0]  [0.1]  [1]    [2]    [3]    [4]    [5]    [6]    [7.0]  [7.1]  [8.0]  [8.1]  [9]    [10]   [11]   [12]   [13]   [14]   [15]   [16.0] [16.1] [17]   [18]
 598 # CHECK-NEXT: 6.00   6.00    -      -      -      -      -      -      -      -     6.00   6.00    -      -     2.00   2.00    -      -      -     6.00   6.00    -      -
 599
 600 # CHECK:      Resource pressure by instruction:
 601 # CHECK-NEXT: [0.0]  [0.1]  [1]    [2]    [3]    [4]    [5]    [6]    [7.0]  [7.1]  [8.0]  [8.1]  [9]    [10]   [11]   [12]   [13]   [14]   [15]   [16.0] [16.1] [17]   [18]   Instructions:
 602 # CHECK-NEXT:  -     3.00    -      -      -      -      -      -      -      -      -     3.00    -      -      -     1.00    -      -      -      -     3.00    -      -     movd     (%rax), %mm0
 603 # CHECK-NEXT: 3.00    -      -      -      -      -      -      -      -      -     3.00    -      -      -     1.00    -      -      -      -     3.00    -      -      -     movd     (%rcx), %mm1
 604 # CHECK-NEXT:  -     3.00    -      -      -      -      -      -      -      -      -     3.00    -      -      -     1.00    -      -      -      -     3.00    -      -     movd     (%rdx), %mm2
 605 # CHECK-NEXT: 3.00    -      -      -      -      -      -      -      -      -     3.00    -      -      -     1.00    -      -      -      -     3.00    -      -      -     movd     (%rbx), %mm3
 606
 607 # CHECK:      Timeline view:
 608 # CHECK-NEXT:                     0
 609 # CHECK-NEXT: Index     0123456789
 610
 611 # CHECK:      [0,0]     DeeeeeER  .   movd      (%rax), %mm0
 612 # CHECK-NEXT: [0,1]     DeeeeeER  .   movd      (%rcx), %mm1
 613 # CHECK-NEXT: [0,2]     D===eeeeeER   movd      (%rdx), %mm2
 614 # CHECK-NEXT: [0,3]     D===eeeeeER   movd      (%rbx), %mm3
 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       movd (%rax), %mm0
 624 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movd (%rcx), %mm1
 625 # CHECK-NEXT: 2.     1     4.0    4.0    0.0       movd (%rdx), %mm2
 626 # CHECK-NEXT: 3.     1     4.0    4.0    0.0       movd (%rbx), %mm3
 627 # CHECK-NEXT:        1     2.5    2.5    0.0       <total>
 628
 629 # CHECK:      [5] Code Region
 630
 631 # CHECK:      Iterations:        100
 632 # CHECK-NEXT: Instructions:      400
 633 # CHECK-NEXT: Total Cycles:      605
 634 # CHECK-NEXT: Total uOps:        400
 635
 636 # CHECK:      Dispatch Width:    4
 637 # CHECK-NEXT: uOps Per Cycle:    0.66
 638 # CHECK-NEXT: IPC:               0.66
 639 # CHECK-NEXT: Block RThroughput: 6.0
 640
 641 # CHECK:      Instruction Info:
 642 # CHECK-NEXT: [1]: #uOps
 643 # CHECK-NEXT: [2]: Latency
 644 # CHECK-NEXT: [3]: RThroughput
 645 # CHECK-NEXT: [4]: MayLoad
 646 # CHECK-NEXT: [5]: MayStore
 647 # CHECK-NEXT: [6]: HasSideEffects (U)
 648
 649 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 650 # CHECK-NEXT:  1      5     1.50    *                   movaps  (%rax), %xmm0
 651 # CHECK-NEXT:  1      5     1.50    *                   movaps  (%rcx), %xmm1
 652 # CHECK-NEXT:  1      5     1.50    *                   movaps  (%rdx), %xmm2
 653 # CHECK-NEXT:  1      5     1.50    *                   movaps  (%rbx), %xmm3
 654
 655 # CHECK:      Dynamic Dispatch Stall Cycles:
 656 # CHECK-NEXT: RAT     - Register unavailable:                      0
 657 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 658 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            0
 659 # CHECK-NEXT: LQ      - Load queue full:                           533  (88.1%)
 660 # CHECK-NEXT: SQ      - Store queue full:                          0
 661 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 662 # CHECK-NEXT: USH     - Uncategorised Structural Hazard:           0
 663
 664 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 665 # CHECK-NEXT: [# dispatched], [# cycles]
 666 # CHECK-NEXT:  0,              416  (68.8%)
 667 # CHECK-NEXT:  2,              178  (29.4%)
 668 # CHECK-NEXT:  4,              11  (1.8%)
 669
 670 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 671 # CHECK-NEXT: [# issued], [# cycles]
 672 # CHECK-NEXT:  0,          405  (66.9%)
 673 # CHECK-NEXT:  2,          200  (33.1%)
 674
 675 # CHECK:      Scheduler's queue usage:
 676 # CHECK-NEXT: [1] Resource name.
 677 # CHECK-NEXT: [2] Average number of used buffer entries.
 678 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 679 # CHECK-NEXT: [4] Total number of buffer entries.
 680
 681 # CHECK:       [1]            [2]        [3]        [4]
 682 # CHECK-NEXT: PdEX             33         36         40
 683 # CHECK-NEXT: PdFPU            33         36         64
 684 # CHECK-NEXT: PdLoad           37         40         40
 685 # CHECK-NEXT: PdStore          0          0          24
 686
 687 # CHECK:      Resources:
 688 # CHECK-NEXT: [0.0] - PdAGLU01
 689 # CHECK-NEXT: [0.1] - PdAGLU01
 690 # CHECK-NEXT: [1]   - PdBranch
 691 # CHECK-NEXT: [2]   - PdCount
 692 # CHECK-NEXT: [3]   - PdDiv
 693 # CHECK-NEXT: [4]   - PdEX0
 694 # CHECK-NEXT: [5]   - PdEX1
 695 # CHECK-NEXT: [6]   - PdFPCVT
 696 # CHECK-NEXT: [7.0] - PdFPFMA
 697 # CHECK-NEXT: [7.1] - PdFPFMA
 698 # CHECK-NEXT: [8.0] - PdFPMAL
 699 # CHECK-NEXT: [8.1] - PdFPMAL
 700 # CHECK-NEXT: [9]   - PdFPMMA
 701 # CHECK-NEXT: [10]  - PdFPSTO
 702 # CHECK-NEXT: [11]  - PdFPU0
 703 # CHECK-NEXT: [12]  - PdFPU1
 704 # CHECK-NEXT: [13]  - PdFPU2
 705 # CHECK-NEXT: [14]  - PdFPU3
 706 # CHECK-NEXT: [15]  - PdFPXBR
 707 # CHECK-NEXT: [16.0] - PdLoad
 708 # CHECK-NEXT: [16.1] - PdLoad
 709 # CHECK-NEXT: [17]  - PdMul
 710 # CHECK-NEXT: [18]  - PdStore
 711
 712 # CHECK:      Resource pressure per iteration:
 713 # CHECK-NEXT: [0.0]  [0.1]  [1]    [2]    [3]    [4]    [5]    [6]    [7.0]  [7.1]  [8.0]  [8.1]  [9]    [10]   [11]   [12]   [13]   [14]   [15]   [16.0] [16.1] [17]   [18]
 714 # CHECK-NEXT: 6.00   6.00    -      -      -      -      -      -     6.00   6.00    -      -      -      -     2.00   2.00    -      -      -     6.00   6.00    -      -
 715
 716 # CHECK:      Resource pressure by instruction:
 717 # CHECK-NEXT: [0.0]  [0.1]  [1]    [2]    [3]    [4]    [5]    [6]    [7.0]  [7.1]  [8.0]  [8.1]  [9]    [10]   [11]   [12]   [13]   [14]   [15]   [16.0] [16.1] [17]   [18]   Instructions:
 718 # CHECK-NEXT:  -     3.00    -      -      -      -      -      -      -     3.00    -      -      -      -      -     1.00    -      -      -      -     3.00    -      -     movaps   (%rax), %xmm0
 719 # CHECK-NEXT: 3.00    -      -      -      -      -      -      -     3.00    -      -      -      -      -     1.00    -      -      -      -     3.00    -      -      -     movaps   (%rcx), %xmm1
 720 # CHECK-NEXT:  -     3.00    -      -      -      -      -      -      -     3.00    -      -      -      -      -     1.00    -      -      -      -     3.00    -      -     movaps   (%rdx), %xmm2
 721 # CHECK-NEXT: 3.00    -      -      -      -      -      -      -     3.00    -      -      -      -      -     1.00    -      -      -      -     3.00    -      -      -     movaps   (%rbx), %xmm3
 722
 723 # CHECK:      Timeline view:
 724 # CHECK-NEXT:                     0
 725 # CHECK-NEXT: Index     0123456789
 726
 727 # CHECK:      [0,0]     DeeeeeER  .   movaps    (%rax), %xmm0
 728 # CHECK-NEXT: [0,1]     DeeeeeER  .   movaps    (%rcx), %xmm1
 729 # CHECK-NEXT: [0,2]     D===eeeeeER   movaps    (%rdx), %xmm2
 730 # CHECK-NEXT: [0,3]     D===eeeeeER   movaps    (%rbx), %xmm3
 731
 732 # CHECK:      Average Wait times (based on the timeline view):
 733 # CHECK-NEXT: [0]: Executions
 734 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 735 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 736 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 737
 738 # CHECK:            [0]    [1]    [2]    [3]
 739 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       movaps       (%rax), %xmm0
 740 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       movaps       (%rcx), %xmm1
 741 # CHECK-NEXT: 2.     1     4.0    4.0    0.0       movaps       (%rdx), %xmm2
 742 # CHECK-NEXT: 3.     1     4.0    4.0    0.0       movaps       (%rbx), %xmm3
 743 # CHECK-NEXT:        1     2.5    2.5    0.0       <total>
 744
 745 # CHECK:      [6] Code Region
 746
 747 # CHECK:      Iterations:        100
 748 # CHECK-NEXT: Instructions:      400
 749 # CHECK-NEXT: Total Cycles:      605
 750 # CHECK-NEXT: Total uOps:        800
 751
 752 # CHECK:      Dispatch Width:    4
 753 # CHECK-NEXT: uOps Per Cycle:    1.32
 754 # CHECK-NEXT: IPC:               0.66
 755 # CHECK-NEXT: Block RThroughput: 6.0
 756
 757 # CHECK:      Instruction Info:
 758 # CHECK-NEXT: [1]: #uOps
 759 # CHECK-NEXT: [2]: Latency
 760 # CHECK-NEXT: [3]: RThroughput
 761 # CHECK-NEXT: [4]: MayLoad
 762 # CHECK-NEXT: [5]: MayStore
 763 # CHECK-NEXT: [6]: HasSideEffects (U)
 764
 765 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 766 # CHECK-NEXT:  2      5     1.50    *                   vmovaps (%rax), %ymm0
 767 # CHECK-NEXT:  2      5     1.50    *                   vmovaps (%rcx), %ymm1
 768 # CHECK-NEXT:  2      5     1.50    *                   vmovaps (%rdx), %ymm2
 769 # CHECK-NEXT:  2      5     1.50    *                   vmovaps (%rbx), %ymm3
 770
 771 # CHECK:      Dynamic Dispatch Stall Cycles:
 772 # CHECK-NEXT: RAT     - Register unavailable:                      0
 773 # CHECK-NEXT: RCU     - Retire tokens unavailable:                 0
 774 # CHECK-NEXT: SCHEDQ  - Scheduler full:                            0
 775 # CHECK-NEXT: LQ      - Load queue full:                           345  (57.0%)
 776 # CHECK-NEXT: SQ      - Store queue full:                          0
 777 # CHECK-NEXT: GROUP   - Static restrictions on the dispatch group: 0
 778 # CHECK-NEXT: USH     - Uncategorised Structural Hazard:           0
 779
 780 # CHECK:      Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
 781 # CHECK-NEXT: [# dispatched], [# cycles]
 782 # CHECK-NEXT:  0,              405  (66.9%)
 783 # CHECK-NEXT:  4,              200  (33.1%)
 784
 785 # CHECK:      Schedulers - number of cycles where we saw N micro opcodes issued:
 786 # CHECK-NEXT: [# issued], [# cycles]
 787 # CHECK-NEXT:  0,          405  (66.9%)
 788 # CHECK-NEXT:  4,          200  (33.1%)
 789
 790 # CHECK:      Scheduler's queue usage:
 791 # CHECK-NEXT: [1] Resource name.
 792 # CHECK-NEXT: [2] Average number of used buffer entries.
 793 # CHECK-NEXT: [3] Maximum number of used buffer entries.
 794 # CHECK-NEXT: [4] Total number of buffer entries.
 795
 796 # CHECK:       [1]            [2]        [3]        [4]
 797 # CHECK-NEXT: PdEX             33         36         40
 798 # CHECK-NEXT: PdFPU            33         36         64
 799 # CHECK-NEXT: PdLoad           36         40         40
 800 # CHECK-NEXT: PdStore          0          0          24
 801
 802 # CHECK:      Resources:
 803 # CHECK-NEXT: [0.0] - PdAGLU01
 804 # CHECK-NEXT: [0.1] - PdAGLU01
 805 # CHECK-NEXT: [1]   - PdBranch
 806 # CHECK-NEXT: [2]   - PdCount
 807 # CHECK-NEXT: [3]   - PdDiv
 808 # CHECK-NEXT: [4]   - PdEX0
 809 # CHECK-NEXT: [5]   - PdEX1
 810 # CHECK-NEXT: [6]   - PdFPCVT
 811 # CHECK-NEXT: [7.0] - PdFPFMA
 812 # CHECK-NEXT: [7.1] - PdFPFMA
 813 # CHECK-NEXT: [8.0] - PdFPMAL
 814 # CHECK-NEXT: [8.1] - PdFPMAL
 815 # CHECK-NEXT: [9]   - PdFPMMA
 816 # CHECK-NEXT: [10]  - PdFPSTO
 817 # CHECK-NEXT: [11]  - PdFPU0
 818 # CHECK-NEXT: [12]  - PdFPU1
 819 # CHECK-NEXT: [13]  - PdFPU2
 820 # CHECK-NEXT: [14]  - PdFPU3
 821 # CHECK-NEXT: [15]  - PdFPXBR
 822 # CHECK-NEXT: [16.0] - PdLoad
 823 # CHECK-NEXT: [16.1] - PdLoad
 824 # CHECK-NEXT: [17]  - PdMul
 825 # CHECK-NEXT: [18]  - PdStore
 826
 827 # CHECK:      Resource pressure per iteration:
 828 # CHECK-NEXT: [0.0]  [0.1]  [1]    [2]    [3]    [4]    [5]    [6]    [7.0]  [7.1]  [8.0]  [8.1]  [9]    [10]   [11]   [12]   [13]   [14]   [15]   [16.0] [16.1] [17]   [18]
 829 # CHECK-NEXT: 6.00   6.00    -      -      -      -      -      -     6.00   6.00    -      -      -      -     2.00   2.00    -      -      -     6.00   6.00    -      -
 830
 831 # CHECK:      Resource pressure by instruction:
 832 # CHECK-NEXT: [0.0]  [0.1]  [1]    [2]    [3]    [4]    [5]    [6]    [7.0]  [7.1]  [8.0]  [8.1]  [9]    [10]   [11]   [12]   [13]   [14]   [15]   [16.0] [16.1] [17]   [18]   Instructions:
 833 # CHECK-NEXT:  -     3.00    -      -      -      -      -      -      -     3.00    -      -      -      -      -     1.00    -      -      -      -     3.00    -      -     vmovaps  (%rax), %ymm0
 834 # CHECK-NEXT: 3.00    -      -      -      -      -      -      -     3.00    -      -      -      -      -     1.00    -      -      -      -     3.00    -      -      -     vmovaps  (%rcx), %ymm1
 835 # CHECK-NEXT:  -     3.00    -      -      -      -      -      -      -     3.00    -      -      -      -      -     1.00    -      -      -      -     3.00    -      -     vmovaps  (%rdx), %ymm2
 836 # CHECK-NEXT: 3.00    -      -      -      -      -      -      -     3.00    -      -      -      -      -     1.00    -      -      -      -     3.00    -      -      -     vmovaps  (%rbx), %ymm3
 837
 838 # CHECK:      Timeline view:
 839 # CHECK-NEXT:                     0
 840 # CHECK-NEXT: Index     0123456789
 841
 842 # CHECK:      [0,0]     DeeeeeER  .   vmovaps   (%rax), %ymm0
 843 # CHECK-NEXT: [0,1]     DeeeeeER  .   vmovaps   (%rcx), %ymm1
 844 # CHECK-NEXT: [0,2]     .D==eeeeeER   vmovaps   (%rdx), %ymm2
 845 # CHECK-NEXT: [0,3]     .D==eeeeeER   vmovaps   (%rbx), %ymm3
 846
 847 # CHECK:      Average Wait times (based on the timeline view):
 848 # CHECK-NEXT: [0]: Executions
 849 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 850 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 851 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 852
 853 # CHECK:            [0]    [1]    [2]    [3]
 854 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       vmovaps      (%rax), %ymm0
 855 # CHECK-NEXT: 1.     1     1.0    1.0    0.0       vmovaps      (%rcx), %ymm1
 856 # CHECK-NEXT: 2.     1     3.0    3.0    0.0       vmovaps      (%rdx), %ymm2
 857 # CHECK-NEXT: 3.     1     3.0    3.0    0.0       vmovaps      (%rbx), %ymm3
 858 # CHECK-NEXT:        1     2.0    2.0    0.0       <total>