Assume a module with the following definitions:
(a) real, allocatable, dimension(:) :: array
or
(b) real, dimension(500) :: array
Both arrays are contiguous as one cannot allocate strides
allocate(array(1:1000:2)) ! invalid
Thus: Only pointers or dummy arguments can have strides.
If one accesses the array in a loop, this can slow down the calculation
considerably.
Test case:
module a
implicit none
real, dimension(500) :: array
! real, allocatable, dimension(:) :: array
contains
subroutine test(x,n)
integer :: n, i
real :: x(n)
! allocate(array(1:500))
do i = 1, n
x(i) = array(i)
end do
end subroutine test
end module a
end
>From the dump difference:
- (*x)[(int8) i + -1] = array[(int8) i + -1];
+ (*x)[(int8) i + -1] = (*(real4[0:] *) array.data)[(int8) i *
array.dim[0].stride + array.offset];
Full example, see:
http://groups.google.com/group/comp.lang.fortran/browse_thread/thread/8fac8fcf6e93fba7/
("Program slowdown when calling function with dynamic arrays" by Paul Hilscher)
--
Summary: gfortran treads known-to-be-contiguous array as having
strides
Product: gcc
Version: 4.3.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: fortran
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: burnus at gcc dot gnu dot org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33753
