test/tools/llvm-mca/X86/BdVer2/zero-idioms-avx-256.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 -timeline -timeline-max-iterations=3 < %s | FileCheck %s
   3
   4 # TODO: Fix the processor resource usage for zero-idiom YMM XOR instructions.
   5 #       Those vector XOR instructions should only consume 1cy of JFPU1 (instead
   6 #       of 2cy).
   7
   8 # LLVM-MCA-BEGIN ZERO-IDIOM-1
   9
  10 vaddps %ymm0, %ymm0, %ymm1
  11 vxorps %ymm1, %ymm1, %ymm1
  12 vblendps $2, %ymm1, %ymm2, %ymm3
  13
  14 # LLVM-MCA-END
  15
  16 # LLVM-MCA-BEGIN ZERO-IDIOM-2
  17
  18 vaddpd %ymm0, %ymm0, %ymm1
  19 vxorpd %ymm1, %ymm1, %ymm1
  20 vblendpd $2, %ymm1, %ymm2, %ymm3
  21
  22 # LLVM-MCA-END
  23
  24 # LLVM-MCA-BEGIN ZERO-IDIOM-3
  25 vaddps %ymm0, %ymm1, %ymm2
  26 vandnps %ymm2, %ymm2, %ymm3
  27 # LLVM-MCA-END
  28
  29 # LLVM-MCA-BEGIN ZERO-IDIOM-4
  30 vaddps %ymm0, %ymm1, %ymm2
  31 vandnps %ymm2, %ymm2, %ymm3
  32 # LLVM-MCA-END
  33
  34 # LLVM-MCA-BEGIN ZERO-IDIOM-5
  35 vperm2f128 $136, %ymm0, %ymm0, %ymm1
  36 vaddps  %ymm1, %ymm1, %ymm0
  37 # LLVM-MCA-END
  38
  39 # CHECK:      [0] Code Region - ZERO-IDIOM-1
  40
  41 # CHECK:      Iterations:        100
  42 # CHECK-NEXT: Instructions:      300
  43 # CHECK-NEXT: Total Cycles:      354
  44 # CHECK-NEXT: Total uOps:        600
  45
  46 # CHECK:      Dispatch Width:    4
  47 # CHECK-NEXT: uOps Per Cycle:    1.69
  48 # CHECK-NEXT: IPC:               0.85
  49 # CHECK-NEXT: Block RThroughput: 3.5
  50
  51 # CHECK:      Instruction Info:
  52 # CHECK-NEXT: [1]: #uOps
  53 # CHECK-NEXT: [2]: Latency
  54 # CHECK-NEXT: [3]: RThroughput
  55 # CHECK-NEXT: [4]: MayLoad
  56 # CHECK-NEXT: [5]: MayStore
  57 # CHECK-NEXT: [6]: HasSideEffects (U)
  58
  59 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
  60 # CHECK-NEXT:  2      5     1.00                        vaddps  %ymm0, %ymm0, %ymm1
  61 # CHECK-NEXT:  2      2     1.00                        vxorps  %ymm1, %ymm1, %ymm1
  62 # CHECK-NEXT:  2      2     1.50                        vblendps        $2, %ymm1, %ymm2, %ymm3
  63
  64 # CHECK:      Resources:
  65 # CHECK-NEXT: [0.0] - PdAGLU01
  66 # CHECK-NEXT: [0.1] - PdAGLU01
  67 # CHECK-NEXT: [1]   - PdBranch
  68 # CHECK-NEXT: [2]   - PdCount
  69 # CHECK-NEXT: [3]   - PdDiv
  70 # CHECK-NEXT: [4]   - PdEX0
  71 # CHECK-NEXT: [5]   - PdEX1
  72 # CHECK-NEXT: [6]   - PdFPCVT
  73 # CHECK-NEXT: [7.0] - PdFPFMA
  74 # CHECK-NEXT: [7.1] - PdFPFMA
  75 # CHECK-NEXT: [8.0] - PdFPMAL
  76 # CHECK-NEXT: [8.1] - PdFPMAL
  77 # CHECK-NEXT: [9]   - PdFPMMA
  78 # CHECK-NEXT: [10]  - PdFPSTO
  79 # CHECK-NEXT: [11]  - PdFPU0
  80 # CHECK-NEXT: [12]  - PdFPU1
  81 # CHECK-NEXT: [13]  - PdFPU2
  82 # CHECK-NEXT: [14]  - PdFPU3
  83 # CHECK-NEXT: [15]  - PdFPXBR
  84 # CHECK-NEXT: [16.0] - PdLoad
  85 # CHECK-NEXT: [16.1] - PdLoad
  86 # CHECK-NEXT: [17]  - PdMul
  87 # CHECK-NEXT: [18]  - PdStore
  88
  89 # CHECK:      Resource pressure per iteration:
  90 # 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]
  91 # CHECK-NEXT:  -      -      -      -      -      -      -      -     3.51   3.49    -      -      -      -     2.00   3.00    -      -      -      -      -      -      -
  92
  93 # CHECK:      Resource pressure by instruction:
  94 # 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:
  95 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.98   1.02    -      -      -      -     1.00    -      -      -      -      -      -      -      -     vaddps   %ymm0, %ymm0, %ymm1
  96 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00   1.00    -      -      -      -     0.98   1.02    -      -      -      -      -      -      -     vxorps   %ymm1, %ymm1, %ymm1
  97 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.53   1.47    -      -      -      -     0.02   1.98    -      -      -      -      -      -      -     vblendps $2, %ymm1, %ymm2, %ymm3
  98
  99 # CHECK:      Timeline view:
 100 # CHECK-NEXT:                     0123
 101 # CHECK-NEXT: Index     0123456789
 102
 103 # CHECK:      [0,0]     DeeeeeER  .  .   vaddps %ymm0, %ymm0, %ymm1
 104 # CHECK-NEXT: [0,1]     DeeE---R  .  .   vxorps %ymm1, %ymm1, %ymm1
 105 # CHECK-NEXT: [0,2]     .D=eeE-R  .  .   vblendps       $2, %ymm1, %ymm2, %ymm3
 106 # CHECK-NEXT: [1,0]     .D=eeeeeER.  .   vaddps %ymm0, %ymm0, %ymm1
 107 # CHECK-NEXT: [1,1]     . D==eeE-R.  .   vxorps %ymm1, %ymm1, %ymm1
 108 # CHECK-NEXT: [1,2]     . D====eeER  .   vblendps       $2, %ymm1, %ymm2, %ymm3
 109 # CHECK-NEXT: [2,0]     .  D==eeeeeER.   vaddps %ymm0, %ymm0, %ymm1
 110 # CHECK-NEXT: [2,1]     .  D====eeE-R.   vxorps %ymm1, %ymm1, %ymm1
 111 # CHECK-NEXT: [2,2]     .   D=====eeER   vblendps       $2, %ymm1, %ymm2, %ymm3
 112
 113 # CHECK:      Average Wait times (based on the timeline view):
 114 # CHECK-NEXT: [0]: Executions
 115 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 116 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 117 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 118
 119 # CHECK:            [0]    [1]    [2]    [3]
 120 # CHECK-NEXT: 0.     3     2.0    2.0    0.0       vaddps       %ymm0, %ymm0, %ymm1
 121 # CHECK-NEXT: 1.     3     3.0    3.0    1.7       vxorps       %ymm1, %ymm1, %ymm1
 122 # CHECK-NEXT: 2.     3     4.3    0.0    0.3       vblendps     $2, %ymm1, %ymm2, %ymm3
 123
 124 # CHECK:      [1] Code Region - ZERO-IDIOM-2
 125
 126 # CHECK:      Iterations:        100
 127 # CHECK-NEXT: Instructions:      300
 128 # CHECK-NEXT: Total Cycles:      354
 129 # CHECK-NEXT: Total uOps:        600
 130
 131 # CHECK:      Dispatch Width:    4
 132 # CHECK-NEXT: uOps Per Cycle:    1.69
 133 # CHECK-NEXT: IPC:               0.85
 134 # CHECK-NEXT: Block RThroughput: 3.5
 135
 136 # CHECK:      Instruction Info:
 137 # CHECK-NEXT: [1]: #uOps
 138 # CHECK-NEXT: [2]: Latency
 139 # CHECK-NEXT: [3]: RThroughput
 140 # CHECK-NEXT: [4]: MayLoad
 141 # CHECK-NEXT: [5]: MayStore
 142 # CHECK-NEXT: [6]: HasSideEffects (U)
 143
 144 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 145 # CHECK-NEXT:  2      5     1.00                        vaddpd  %ymm0, %ymm0, %ymm1
 146 # CHECK-NEXT:  2      2     1.00                        vxorpd  %ymm1, %ymm1, %ymm1
 147 # CHECK-NEXT:  2      2     1.50                        vblendpd        $2, %ymm1, %ymm2, %ymm3
 148
 149 # CHECK:      Resources:
 150 # CHECK-NEXT: [0.0] - PdAGLU01
 151 # CHECK-NEXT: [0.1] - PdAGLU01
 152 # CHECK-NEXT: [1]   - PdBranch
 153 # CHECK-NEXT: [2]   - PdCount
 154 # CHECK-NEXT: [3]   - PdDiv
 155 # CHECK-NEXT: [4]   - PdEX0
 156 # CHECK-NEXT: [5]   - PdEX1
 157 # CHECK-NEXT: [6]   - PdFPCVT
 158 # CHECK-NEXT: [7.0] - PdFPFMA
 159 # CHECK-NEXT: [7.1] - PdFPFMA
 160 # CHECK-NEXT: [8.0] - PdFPMAL
 161 # CHECK-NEXT: [8.1] - PdFPMAL
 162 # CHECK-NEXT: [9]   - PdFPMMA
 163 # CHECK-NEXT: [10]  - PdFPSTO
 164 # CHECK-NEXT: [11]  - PdFPU0
 165 # CHECK-NEXT: [12]  - PdFPU1
 166 # CHECK-NEXT: [13]  - PdFPU2
 167 # CHECK-NEXT: [14]  - PdFPU3
 168 # CHECK-NEXT: [15]  - PdFPXBR
 169 # CHECK-NEXT: [16.0] - PdLoad
 170 # CHECK-NEXT: [16.1] - PdLoad
 171 # CHECK-NEXT: [17]  - PdMul
 172 # CHECK-NEXT: [18]  - PdStore
 173
 174 # CHECK:      Resource pressure per iteration:
 175 # 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]
 176 # CHECK-NEXT:  -      -      -      -      -      -      -      -     3.51   3.49    -      -      -      -     2.00   3.00    -      -      -      -      -      -      -
 177
 178 # CHECK:      Resource pressure by instruction:
 179 # 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:
 180 # CHECK-NEXT:  -      -      -      -      -      -      -      -     0.98   1.02    -      -      -      -     1.00    -      -      -      -      -      -      -      -     vaddpd   %ymm0, %ymm0, %ymm1
 181 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.00   1.00    -      -      -      -     0.98   1.02    -      -      -      -      -      -      -     vxorpd   %ymm1, %ymm1, %ymm1
 182 # CHECK-NEXT:  -      -      -      -      -      -      -      -     1.53   1.47    -      -      -      -     0.02   1.98    -      -      -      -      -      -      -     vblendpd $2, %ymm1, %ymm2, %ymm3
 183
 184 # CHECK:      Timeline view:
 185 # CHECK-NEXT:                     0123
 186 # CHECK-NEXT: Index     0123456789
 187
 188 # CHECK:      [0,0]     DeeeeeER  .  .   vaddpd %ymm0, %ymm0, %ymm1
 189 # CHECK-NEXT: [0,1]     DeeE---R  .  .   vxorpd %ymm1, %ymm1, %ymm1
 190 # CHECK-NEXT: [0,2]     .D=eeE-R  .  .   vblendpd       $2, %ymm1, %ymm2, %ymm3
 191 # CHECK-NEXT: [1,0]     .D=eeeeeER.  .   vaddpd %ymm0, %ymm0, %ymm1
 192 # CHECK-NEXT: [1,1]     . D==eeE-R.  .   vxorpd %ymm1, %ymm1, %ymm1
 193 # CHECK-NEXT: [1,2]     . D====eeER  .   vblendpd       $2, %ymm1, %ymm2, %ymm3
 194 # CHECK-NEXT: [2,0]     .  D==eeeeeER.   vaddpd %ymm0, %ymm0, %ymm1
 195 # CHECK-NEXT: [2,1]     .  D====eeE-R.   vxorpd %ymm1, %ymm1, %ymm1
 196 # CHECK-NEXT: [2,2]     .   D=====eeER   vblendpd       $2, %ymm1, %ymm2, %ymm3
 197
 198 # CHECK:      Average Wait times (based on the timeline view):
 199 # CHECK-NEXT: [0]: Executions
 200 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 201 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 202 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 203
 204 # CHECK:            [0]    [1]    [2]    [3]
 205 # CHECK-NEXT: 0.     3     2.0    2.0    0.0       vaddpd       %ymm0, %ymm0, %ymm1
 206 # CHECK-NEXT: 1.     3     3.0    3.0    1.7       vxorpd       %ymm1, %ymm1, %ymm1
 207 # CHECK-NEXT: 2.     3     4.3    0.0    0.3       vblendpd     $2, %ymm1, %ymm2, %ymm3
 208
 209 # CHECK:      [2] Code Region - ZERO-IDIOM-3
 210
 211 # CHECK:      Iterations:        100
 212 # CHECK-NEXT: Instructions:      200
 213 # CHECK-NEXT: Total Cycles:      206
 214 # CHECK-NEXT: Total uOps:        400
 215
 216 # CHECK:      Dispatch Width:    4
 217 # CHECK-NEXT: uOps Per Cycle:    1.94
 218 # CHECK-NEXT: IPC:               0.97
 219 # CHECK-NEXT: Block RThroughput: 2.0
 220
 221 # CHECK:      Instruction Info:
 222 # CHECK-NEXT: [1]: #uOps
 223 # CHECK-NEXT: [2]: Latency
 224 # CHECK-NEXT: [3]: RThroughput
 225 # CHECK-NEXT: [4]: MayLoad
 226 # CHECK-NEXT: [5]: MayStore
 227 # CHECK-NEXT: [6]: HasSideEffects (U)
 228
 229 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 230 # CHECK-NEXT:  2      5     1.00                        vaddps  %ymm0, %ymm1, %ymm2
 231 # CHECK-NEXT:  2      2     1.00                        vandnps %ymm2, %ymm2, %ymm3
 232
 233 # CHECK:      Resources:
 234 # CHECK-NEXT: [0.0] - PdAGLU01
 235 # CHECK-NEXT: [0.1] - PdAGLU01
 236 # CHECK-NEXT: [1]   - PdBranch
 237 # CHECK-NEXT: [2]   - PdCount
 238 # CHECK-NEXT: [3]   - PdDiv
 239 # CHECK-NEXT: [4]   - PdEX0
 240 # CHECK-NEXT: [5]   - PdEX1
 241 # CHECK-NEXT: [6]   - PdFPCVT
 242 # CHECK-NEXT: [7.0] - PdFPFMA
 243 # CHECK-NEXT: [7.1] - PdFPFMA
 244 # CHECK-NEXT: [8.0] - PdFPMAL
 245 # CHECK-NEXT: [8.1] - PdFPMAL
 246 # CHECK-NEXT: [9]   - PdFPMMA
 247 # CHECK-NEXT: [10]  - PdFPSTO
 248 # CHECK-NEXT: [11]  - PdFPU0
 249 # CHECK-NEXT: [12]  - PdFPU1
 250 # CHECK-NEXT: [13]  - PdFPU2
 251 # CHECK-NEXT: [14]  - PdFPU3
 252 # CHECK-NEXT: [15]  - PdFPXBR
 253 # CHECK-NEXT: [16.0] - PdLoad
 254 # CHECK-NEXT: [16.1] - PdLoad
 255 # CHECK-NEXT: [17]  - PdMul
 256 # CHECK-NEXT: [18]  - PdStore
 257
 258 # CHECK:      Resource pressure per iteration:
 259 # 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]
 260 # CHECK-NEXT:  -      -      -      -      -      -      -      -     2.00   2.00    -      -      -      -     1.00   2.00    -      -      -      -      -      -      -
 261
 262 # CHECK:      Resource pressure by instruction:
 263 # 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:
 264 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     2.00    -      -      -      -     1.00    -      -      -      -      -      -      -      -     vaddps   %ymm0, %ymm1, %ymm2
 265 # CHECK-NEXT:  -      -      -      -      -      -      -      -     2.00    -      -      -      -      -      -     2.00    -      -      -      -      -      -      -     vandnps  %ymm2, %ymm2, %ymm3
 266
 267 # CHECK:      Timeline view:
 268 # CHECK-NEXT:                     01
 269 # CHECK-NEXT: Index     0123456789
 270
 271 # CHECK:      [0,0]     DeeeeeER  ..   vaddps   %ymm0, %ymm1, %ymm2
 272 # CHECK-NEXT: [0,1]     DeeE---R  ..   vandnps  %ymm2, %ymm2, %ymm3
 273 # CHECK-NEXT: [1,0]     .D=eeeeeER..   vaddps   %ymm0, %ymm1, %ymm2
 274 # CHECK-NEXT: [1,1]     .D=eeE---R..   vandnps  %ymm2, %ymm2, %ymm3
 275 # CHECK-NEXT: [2,0]     . D==eeeeeER   vaddps   %ymm0, %ymm1, %ymm2
 276 # CHECK-NEXT: [2,1]     . D==eeE---R   vandnps  %ymm2, %ymm2, %ymm3
 277
 278 # CHECK:      Average Wait times (based on the timeline view):
 279 # CHECK-NEXT: [0]: Executions
 280 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 281 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 282 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 283
 284 # CHECK:            [0]    [1]    [2]    [3]
 285 # CHECK-NEXT: 0.     3     2.0    2.0    0.0       vaddps       %ymm0, %ymm1, %ymm2
 286 # CHECK-NEXT: 1.     3     2.0    2.0    3.0       vandnps      %ymm2, %ymm2, %ymm3
 287
 288 # CHECK:      [3] Code Region - ZERO-IDIOM-4
 289
 290 # CHECK:      Iterations:        100
 291 # CHECK-NEXT: Instructions:      200
 292 # CHECK-NEXT: Total Cycles:      206
 293 # CHECK-NEXT: Total uOps:        400
 294
 295 # CHECK:      Dispatch Width:    4
 296 # CHECK-NEXT: uOps Per Cycle:    1.94
 297 # CHECK-NEXT: IPC:               0.97
 298 # CHECK-NEXT: Block RThroughput: 2.0
 299
 300 # CHECK:      Instruction Info:
 301 # CHECK-NEXT: [1]: #uOps
 302 # CHECK-NEXT: [2]: Latency
 303 # CHECK-NEXT: [3]: RThroughput
 304 # CHECK-NEXT: [4]: MayLoad
 305 # CHECK-NEXT: [5]: MayStore
 306 # CHECK-NEXT: [6]: HasSideEffects (U)
 307
 308 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 309 # CHECK-NEXT:  2      5     1.00                        vaddps  %ymm0, %ymm1, %ymm2
 310 # CHECK-NEXT:  2      2     1.00                        vandnps %ymm2, %ymm2, %ymm3
 311
 312 # CHECK:      Resources:
 313 # CHECK-NEXT: [0.0] - PdAGLU01
 314 # CHECK-NEXT: [0.1] - PdAGLU01
 315 # CHECK-NEXT: [1]   - PdBranch
 316 # CHECK-NEXT: [2]   - PdCount
 317 # CHECK-NEXT: [3]   - PdDiv
 318 # CHECK-NEXT: [4]   - PdEX0
 319 # CHECK-NEXT: [5]   - PdEX1
 320 # CHECK-NEXT: [6]   - PdFPCVT
 321 # CHECK-NEXT: [7.0] - PdFPFMA
 322 # CHECK-NEXT: [7.1] - PdFPFMA
 323 # CHECK-NEXT: [8.0] - PdFPMAL
 324 # CHECK-NEXT: [8.1] - PdFPMAL
 325 # CHECK-NEXT: [9]   - PdFPMMA
 326 # CHECK-NEXT: [10]  - PdFPSTO
 327 # CHECK-NEXT: [11]  - PdFPU0
 328 # CHECK-NEXT: [12]  - PdFPU1
 329 # CHECK-NEXT: [13]  - PdFPU2
 330 # CHECK-NEXT: [14]  - PdFPU3
 331 # CHECK-NEXT: [15]  - PdFPXBR
 332 # CHECK-NEXT: [16.0] - PdLoad
 333 # CHECK-NEXT: [16.1] - PdLoad
 334 # CHECK-NEXT: [17]  - PdMul
 335 # CHECK-NEXT: [18]  - PdStore
 336
 337 # CHECK:      Resource pressure per iteration:
 338 # 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]
 339 # CHECK-NEXT:  -      -      -      -      -      -      -      -     2.00   2.00    -      -      -      -     1.00   2.00    -      -      -      -      -      -      -
 340
 341 # CHECK:      Resource pressure by instruction:
 342 # 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:
 343 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     2.00    -      -      -      -     1.00    -      -      -      -      -      -      -      -     vaddps   %ymm0, %ymm1, %ymm2
 344 # CHECK-NEXT:  -      -      -      -      -      -      -      -     2.00    -      -      -      -      -      -     2.00    -      -      -      -      -      -      -     vandnps  %ymm2, %ymm2, %ymm3
 345
 346 # CHECK:      Timeline view:
 347 # CHECK-NEXT:                     01
 348 # CHECK-NEXT: Index     0123456789
 349
 350 # CHECK:      [0,0]     DeeeeeER  ..   vaddps   %ymm0, %ymm1, %ymm2
 351 # CHECK-NEXT: [0,1]     DeeE---R  ..   vandnps  %ymm2, %ymm2, %ymm3
 352 # CHECK-NEXT: [1,0]     .D=eeeeeER..   vaddps   %ymm0, %ymm1, %ymm2
 353 # CHECK-NEXT: [1,1]     .D=eeE---R..   vandnps  %ymm2, %ymm2, %ymm3
 354 # CHECK-NEXT: [2,0]     . D==eeeeeER   vaddps   %ymm0, %ymm1, %ymm2
 355 # CHECK-NEXT: [2,1]     . D==eeE---R   vandnps  %ymm2, %ymm2, %ymm3
 356
 357 # CHECK:      Average Wait times (based on the timeline view):
 358 # CHECK-NEXT: [0]: Executions
 359 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 360 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 361 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 362
 363 # CHECK:            [0]    [1]    [2]    [3]
 364 # CHECK-NEXT: 0.     3     2.0    2.0    0.0       vaddps       %ymm0, %ymm1, %ymm2
 365 # CHECK-NEXT: 1.     3     2.0    2.0    3.0       vandnps      %ymm2, %ymm2, %ymm3
 366
 367 # CHECK:      [4] Code Region - ZERO-IDIOM-5
 368
 369 # CHECK:      Iterations:        100
 370 # CHECK-NEXT: Instructions:      200
 371 # CHECK-NEXT: Total Cycles:      903
 372 # CHECK-NEXT: Total uOps:        1000
 373
 374 # CHECK:      Dispatch Width:    4
 375 # CHECK-NEXT: uOps Per Cycle:    1.11
 376 # CHECK-NEXT: IPC:               0.22
 377 # CHECK-NEXT: Block RThroughput: 4.0
 378
 379 # CHECK:      Instruction Info:
 380 # CHECK-NEXT: [1]: #uOps
 381 # CHECK-NEXT: [2]: Latency
 382 # CHECK-NEXT: [3]: RThroughput
 383 # CHECK-NEXT: [4]: MayLoad
 384 # CHECK-NEXT: [5]: MayStore
 385 # CHECK-NEXT: [6]: HasSideEffects (U)
 386
 387 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 388 # CHECK-NEXT:  8      4     3.00                        vperm2f128      $136, %ymm0, %ymm0, %ymm1
 389 # CHECK-NEXT:  2      5     1.00                        vaddps  %ymm1, %ymm1, %ymm0
 390
 391 # CHECK:      Resources:
 392 # CHECK-NEXT: [0.0] - PdAGLU01
 393 # CHECK-NEXT: [0.1] - PdAGLU01
 394 # CHECK-NEXT: [1]   - PdBranch
 395 # CHECK-NEXT: [2]   - PdCount
 396 # CHECK-NEXT: [3]   - PdDiv
 397 # CHECK-NEXT: [4]   - PdEX0
 398 # CHECK-NEXT: [5]   - PdEX1
 399 # CHECK-NEXT: [6]   - PdFPCVT
 400 # CHECK-NEXT: [7.0] - PdFPFMA
 401 # CHECK-NEXT: [7.1] - PdFPFMA
 402 # CHECK-NEXT: [8.0] - PdFPMAL
 403 # CHECK-NEXT: [8.1] - PdFPMAL
 404 # CHECK-NEXT: [9]   - PdFPMMA
 405 # CHECK-NEXT: [10]  - PdFPSTO
 406 # CHECK-NEXT: [11]  - PdFPU0
 407 # CHECK-NEXT: [12]  - PdFPU1
 408 # CHECK-NEXT: [13]  - PdFPU2
 409 # CHECK-NEXT: [14]  - PdFPU3
 410 # CHECK-NEXT: [15]  - PdFPXBR
 411 # CHECK-NEXT: [16.0] - PdLoad
 412 # CHECK-NEXT: [16.1] - PdLoad
 413 # CHECK-NEXT: [17]  - PdMul
 414 # CHECK-NEXT: [18]  - PdStore
 415
 416 # CHECK:      Resource pressure per iteration:
 417 # 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]
 418 # CHECK-NEXT:  -      -      -      -      -      -      -      -     2.00   6.00    -      -      -      -     1.00   1.00    -      -      -      -      -      -      -
 419
 420 # CHECK:      Resource pressure by instruction:
 421 # 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:
 422 # CHECK-NEXT:  -      -      -      -      -      -      -      -      -     6.00    -      -      -      -      -     1.00    -      -      -      -      -      -      -     vperm2f128       $136, %ymm0, %ymm0, %ymm1
 423 # CHECK-NEXT:  -      -      -      -      -      -      -      -     2.00    -      -      -      -      -     1.00    -      -      -      -      -      -      -      -     vaddps   %ymm1, %ymm1, %ymm0
 424
 425 # CHECK:      Timeline view:
 426 # CHECK-NEXT:                     0123456789
 427 # CHECK-NEXT: Index     0123456789          0123456789
 428
 429 # CHECK:      [0,0]     DeeeeER   .    .    .    .   .   vperm2f128     $136, %ymm0, %ymm0, %ymm1
 430 # CHECK-NEXT: [0,1]     . D==eeeeeER   .    .    .   .   vaddps %ymm1, %ymm1, %ymm0
 431 # CHECK-NEXT: [1,0]     .  D======eeeeER    .    .   .   vperm2f128     $136, %ymm0, %ymm0, %ymm1
 432 # CHECK-NEXT: [1,1]     .    D========eeeeeER    .   .   vaddps %ymm1, %ymm1, %ymm0
 433 # CHECK-NEXT: [2,0]     .    .D============eeeeER.   .   vperm2f128     $136, %ymm0, %ymm0, %ymm1
 434 # CHECK-NEXT: [2,1]     .    .  D==============eeeeeER   vaddps %ymm1, %ymm1, %ymm0
 435
 436 # CHECK:      Average Wait times (based on the timeline view):
 437 # CHECK-NEXT: [0]: Executions
 438 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
 439 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
 440 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
 441
 442 # CHECK:            [0]    [1]    [2]    [3]
 443 # CHECK-NEXT: 0.     3     7.0    0.3    0.0       vperm2f128   $136, %ymm0, %ymm0, %ymm1
 444 # CHECK-NEXT: 1.     3     9.0    0.0    0.0       vaddps       %ymm1, %ymm1, %ymm0