[PATCH] D30431: [PowerPC] MachineSSA pass to reduce the number of CR-logical operations

[Nemanja Ivanovic via Phabricator via llvm-commits]

nemanjai added a comment.

Some performance numbers as requested by @echristo (all run on a Power8 2GHz machine, bound to a specific physical CPU):
SPEC2006 (run time improvements - negative is good):
444.namd,868.7807,0.0000,850.4903,0.0000,-2.11%
447.dealII,630.6040,0.0000,631.2014,0.0000,0.09%
450.soplex,473.6103,0.0000,417.0427,0.0000,-11.94%
453.povray,505.4464,0.0000,487.0534,0.0000,-3.64%
401.bzip2,969.0608,0.0000,923.7337,0.0000,-4.68%
445.gobmk,868.6733,0.0000,857.6179,0.0000,-1.27%
464.h264ref,1053.8501,0.0000,1005.3811,0.0000,-4.60%
471.omnetpp,473.8660,0.0000,444.2557,0.0000,-6.25%
Changes below 1% omitted.
This was from a single run of SPEC2016 with r319300 and then same revision with this patch applied. Since it was a single run, I don't have information about the noise range, but the important thing is that the trend seems to be rather clear - performance improves with this patch.

For Hal's benchmarks, the relative performance doesn't change almost at all with this patch. The absolute performance improves with `4,3,float`, `4,4,float` by about 30% and degrades with `4,4,int` by about 20%. The rest are unchanged.


Repository:
  rL LLVM

https://reviews.llvm.org/D30431