[PATCH] [MachineSink] Use the real post dominator tree
Jingyue Wu
jingyue at google.com
Mon Oct 6 15:32:14 PDT 2014
Runtime performance per SPEC CINT2006. The four columns are benchmark name, reference time (not useful here), run time, and ratio vs reference (not useful here).
Original
```
400.perlbench 9770 320 30.5 S
400.perlbench 9770 323 30.3 *
400.perlbench 9770 324 30.2 S
401.bzip2 9650 420 23.0 S
401.bzip2 9650 451 21.4 S
401.bzip2 9650 420 23.0 *
403.gcc 8050 249 32.3 S
403.gcc 8050 256 31.5 S
403.gcc 8050 250 32.1 *
429.mcf 9120 280 32.6 *
429.mcf 9120 283 32.2 S
429.mcf 9120 279 32.7 S
445.gobmk 10490 407 25.8 *
445.gobmk 10490 407 25.8 S
445.gobmk 10490 407 25.8 S
456.hmmer 9330 363 25.7 S
456.hmmer 9330 363 25.7 *
456.hmmer 9330 365 25.6 S
458.sjeng 12100 492 24.6 S
458.sjeng 12100 490 24.7 *
458.sjeng 12100 489 24.7 S
462.libquantum 20720 396 52.4 S
462.libquantum 20720 403 51.4 *
462.libquantum 20720 450 46.0 S
464.h264ref 22130 460 48.1 *
464.h264ref 22130 460 48.1 S
464.h264ref 22130 461 48.0 S
471.omnetpp 6250 230 27.2 *
471.omnetpp 6250 231 27.1 S
471.omnetpp 6250 229 27.3 S
473.astar 7020 336 20.9 *
473.astar 7020 338 20.8 S
473.astar 7020 336 20.9 S
483.xalancbmk 6900 186 37.1 *
483.xalancbmk 6900 196 35.2 S
483.xalancbmk 6900 185 37.2 S
```
With this change
```
400.perlbench 9770 307 31.9 S
400.perlbench 9770 307 31.8 S
400.perlbench 9770 307 31.9 *
401.bzip2 9650 419 23.0 *
401.bzip2 9650 419 23.0 S
401.bzip2 9650 419 23.0 S
403.gcc 8050 251 32.0 S
403.gcc 8050 249 32.3 S
403.gcc 8050 251 32.1 *
429.mcf 9120 273 33.4 S
429.mcf 9120 280 32.6 S
429.mcf 9120 276 33.1 *
445.gobmk 10490 406 25.9 S
445.gobmk 10490 406 25.8 *
445.gobmk 10490 410 25.6 S
456.hmmer 9330 360 25.9 S
456.hmmer 9330 361 25.9 *
456.hmmer 9330 361 25.8 S
458.sjeng 12100 483 25.0 S
458.sjeng 12100 485 24.9 *
458.sjeng 12100 486 24.9 S
462.libquantum 20720 364 56.9 S
462.libquantum 20720 376 55.2 *
462.libquantum 20720 381 54.3 S
464.h264ref 22130 457 48.4 S
464.h264ref 22130 456 48.5 *
464.h264ref 22130 455 48.6 S
471.omnetpp 6250 228 27.4 S
471.omnetpp 6250 229 27.3 S
471.omnetpp 6250 228 27.4 *
473.astar 7020 335 21.0 S
473.astar 7020 333 21.1 S
473.astar 7020 333 21.1 *
483.xalancbmk 6900 186 37.2 *
483.xalancbmk 6900 185 37.4 S
483.xalancbmk 6900 186 37.1 S
```
Each benchmark was run three times and the results look stable across runs. We take the median and compute the slowdown/speedup.
```
benchmark original sink slowdown
400.perlbench 323 307 -4.953560372
401.bzip2 420 419 -0.2380952381
403.gcc 250 251 0.4
429.mcf 280 276 -1.428571429
445.gobmk 407 406 -0.2457002457
456.hmmer 363 361 -0.5509641873
458.sjeng 490 485 -1.020408163
462.libquantum 403 376 -6.699751861
464.h264ref 460 456 -0.8695652174
471.omnetpp 230 228 -0.8695652174
473.astar 336 333 -0.8928571429
483.xalancbmk 186 186 0
```
We see significant speedup on 400.perlbench and 462.libquantum without any significant regression.
http://reviews.llvm.org/D5633
More information about the llvm-commits
mailing list