Hello,
I encountered what I consider a compiler bug. Unfortunatelly, during the past 30min I was not able to get "premier support" to work or find some other way of submitting this issue more directly.
Concretely:
The following file
-- bug.f90 --
module dummy_mod
contains
subroutine dummyfn(nOcc)
integer, intent(in) ::&
nOcc
integer ::&
DiagonalPairs(2, nOcc)
DiagonalPairs(1,:) = (/ (i, i = 1, nOcc) /)
DiagonalPairs(2,:) = DiagonalPairs(1,:)
! ^- second statement produces broken results on ifort
! 10.1 20070913 if compiled with -fPIC -i8 -O3 bug.f90
do i = 1, nOcc
print *, DiagonalPairs(1,i), DiagonalPairs(2,i)
end do
end subroutine
end module
program broken_array_assign
use dummy_mod
call dummyfn(5)
end program
-- eof --
does as expected if compiled with "ifort -fPIC -i8 -O0 bug.f90", but breaks if compiled with "ifort -fPIC -i8 -O3 bug.f90":
~/tmp/ifort10.1-bug> run.sh
with 'ifort -fPIC -i8 -O0 bug.f90':
1 1
2 2
3 3
4 4
5 5
with 'ifort -fPIC -i8 -O3 bug.f90':
bug.f90(11): (col. 4) remark: FUSED LOOP WAS VECTORIZED.
1 0
2 0
3 72105053915740032
4 47459878089888
&n
bsp; 5 5
note: this does not happen without the -fPIC or in 32bit mode.
~/tmp/ifort10.1-bug>
I'd appreciate it if someone from Intel reading this forum would forward this to the suitable compiler developers, as I can't do it.
