https://gcc.gnu.org/bugzilla/show_bug.cgi?id=27855
Aldy Hernandez <aldyh at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |aldyh at gcc dot gnu.org
--- Comment #54 from Aldy Hernandez <aldyh at gcc dot gnu.org> ---
Results from comparing trunk and gcc-4_6-branch on (average of 7 runs):
Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz
4.6.x
-O2 -ffast-math -msse4.1 -DTYPE=double -fno-tree-reassoc
atlasmm 60 1000 0.100 5415.4
-O2 -ffast-math -msse4.1 -DTYPE=double
atlasmm 60 1000 0.082 4797.4
Performance penalty: 12.9%
trunk:
-O2 -ffast-math -msse4.1 -DTYPE=double -fno-tree-reassoc
atlasmm 60 1000 0.105 5217.0
-O2 -ffast-math -msse4.1 -DTYPE=double
atlasmm 60 1000 0.108 4773.0
Performance change: 9.3%
It looks like the penalty for using fno-tree-reassoc now is even less than it
was in 4.6.x. The original report was for a 30% drop.
Has this been fixed, or am I missing something? (different flags for my 4.6
run? Is the 4.6 branch not a suitable representation of the 4.6.2 compiler
against which this bug was originally filed against?)
