http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60646
kargl at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |kargl at gcc dot gnu.org
--- Comment #1 from kargl at gcc dot gnu.org ---
Janne,
Can you be a little more specific on what you want
to investigate? AFAIK, the general handling of
complex division is done be the middle-end.
% cat l.f90
subroutine loo(z1,z2,z3)
complex z1, z2, z3
z3 = z1 / z2
end subroutine loo
% gfc4x -fdump-tree-all -c l.f90
% ls l.f90.*
l.f90.003t.original l.f90.018t.einline
l.f90.004t.gimple l.f90.033t.profile_estimate
l.f90.006t.omplower l.f90.036t.release_ssa
l.f90.007t.lower l.f90.037t.inline_param2
l.f90.010t.eh l.f90.161t.veclower
l.f90.011t.cfg l.f90.162t.cplxlower0
l.f90.015t.ssa l.f90.169t.optimized
l.f90.017t.inline_param1 l.f90.249t.statistics
l.f90.cplxlower0 is definitely coming from the
middle-end.
The Fortran front-end will perform constant-folding.
% cat k.f90
subroutine lu(z)
complex z
complex, parameter :: a = (1,2), b=(3,4)
z = a/b
end subroutine lu
% cat k.f90.003t.original
lu (complex(kind=4) & restrict z)
{
*z=__complex__(4.39999997615814208984375e-1,7.999999821186065673828125e-2);
}
Note, the front-end takes advantage of MPFR to do the folding, and so
it should yield an accurate result.
