https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64425
--- Comment #3 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
Yep, I looked into this before comitting the patch. I have tracked it down to
inlining of function UniformRectilinearMesh<MeshTraits<3, double,
UniformRectilinearTag, CartesianTag, 3> >::cellPosition(Loc<3> const&) const:
inline PointType_t cellPosition(const Loc_t &loc) const
{
PointType_t point;
for (int i=0; i<dimensions; i++)
point(i) = origin()(i) + spacings()(i)
* (loc[i].first() - physicalCellDomain()[i].first() + 0.5);
return point;
}
this function is somewhat evil: dimensons is a small compile time constant and
the loop unrolls completely into quite simple code. This is however not done
prior inlining. Inliner thus think that the function is 168 instructions and
the speedup for inlining accounts only removing some of this-> references,
overall 9%.
As an effect of various roundoff errors being eliminated by the sreal code, the
functions sinks down in the priority queue quite a bit bellow the
inline-unit-growth cutoff. There seems to be no issues with priority
calculation, just the functions looks bit uncool for inlining and got lucky in
the old model (that was for years tested on tramp3d)
The patch I sent for simplification of priority queue solves the failure:
https://gcc.gnu.org/ml/gcc-patches/2015-01/msg00004.html
The patch seem win in general on benchmarks I tested (it is big win for tramp3d
so I can reduce unit growth to 15%) I however would like first the testers to
get restarted after vacations and there is one regression I want to analyze
first in my testing.
