------- Comment #6 from burnus at gcc dot gnu dot org 2006-12-01 00:04 -------
> You misunderstood me.
> const int NODES = 2500;
> float f[NODES];
> causes a VLA to happen because in C, NODES is not a constant expression.
> You can do instead:
> #define NODES 2500
> to get the non VLA.
I should really read a good C book.
gcc: 0m1.390s
icc: 0m2.439s
Thus converting such VLA arrays into normal arrays seems to be indeed what icc
does -- and GCC should do!?
The next question is of cause whether this has anything to do with my Fortran
slowness. Using oprofile, I found with for gfortran in gas_dyn.f90:
150879 51.8663 : CS(:NODES) = SQRT(CGAMMA*PRES(:NODES)/DENS(:NODES))
In any case, for the following program I get:
gfortran -O4 -msse3 -ftree-vectorize -funroll-loops -march=opteron -ffast-math
0m7.281s
ifort -O3 -xW
0m2.633s
program main
implicit none
integer, parameter :: NODES = 25000
real :: CGAMMA
real :: DENS(NODES), CS(NODES), PRES(NODES)
integer :: i
CGAMMA = sqrt(2.0)
DENS = sqrt(3.0)
PRES = sin(0.25)
CGAMMA = 3.0
do i=1,20000
call EOS(NODES,CGAMMA, DENS,PRES,CS)
CGAMMA = CGAMMA + CS(1)
end do
print *,CS(1)
contains
SUBROUTINE EOS(NODES, CGAMMA, DENS, PRES, CS)
IMPLICIT NONE
integer :: i
INTEGER NODES
REAL CGAMMA
REAL, DIMENSION(NODES) :: DENS, PRES, CS
do i = 1,NODES
CS(i) = sqrt(cgamma*PRES(i)/Dens(i))
end do
end subroutine
end program main
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30032
