[PATCH] D115406: [SystemZ] Improve eliminateFrameIndex() to use STDY/STEY/LDY for VR32/VR64 when possible.

[Jonas Paulsson via Phabricator via llvm-commits]

jonpa added a comment.

I made a new set of quick benchmark runs with fp-contract=default/off/fast for main/patch/"patch with LEYs", giving 9 runs for each benchmark.

On Z14
default: cactus improving 1.5%, LEY:s give 1.2% improvement on blender (no regressions)
fast: cactus improving 2.5%, blender improving 0.5%. LEY:s give no further improvements (but possible slight regression on xalank 0.4%).
off: cactus improving 2%. LEY:s give 1.1% improvement on blender (no regressions).

On Z15
default: cactus improving 1%, nab improving 0.7%, deepsheng regressing 0.8%. Imagick regressing 2.3%, also without LEYs. LEYs give 1% improvement on perl.
fast: cactus improving 1.8%
off: cactus improving 2.4%. LEYs improve perl 1%, but cause 5% regression on imagick.

Z15/default/Imagick regresssion 2.3%:

- seems to be <1% with a "full" run (~0.7%)
- Opcode diff look ok:

  lay            :                 5776                 5449     -327
  ld             :                 4160                 4004     -156
  ldy            :                  684                  840     +156
  stey           :                   36                  138     +102
  ste            :                 1776                 1674     -102
  std            :                 2786                 2713      -73
  stdy           :                  550                  623      +73
  lg             :                28283                28275       -8
  stg            :                20219                20211       -8

Hot file is morphology.s:

  lay            :                  221                  120     -101
  stey           :                    0                   93      +93
  ste            :                   93                    0      -93
  stg            :                  293                  285       -8
  lg             :                  633                  625       -8
  std            :                  224                  219       -5
  stdy           :                   14                   19       +5
  ld             :                  316                  312       -4
  ldy            :                   31                   35       +4

Looks ok as well.. (Less STG/LG are cases where the LAYs needed one extra register.) Not sure then what this could be, or if it's worth more time...


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D115406/new/

https://reviews.llvm.org/D115406