Follow up to PR 41907.
----------------
module m
contains
subroutine one(a)
real, dimension(:,:), intent(inout) :: a
call two(a)
end subroutine one
subroutine two(a)
real, dimension(:,:), intent(inout) :: a
end subroutine two
end module
----------------
As both "one" and "two" have a assumed-shape argument (deferred shape would be
the same), there is absolutely no reason to copy the array descriptor. But
gfortran does so:
a.0 = (real(kind=4)[0:D.1405] * restrict) a->data;
parm.6.data = (void *) &(*a.0)[0];
two (&parm.6);
For assume-size/explicit-size arrays, gfortran does do it correctly. Also for
deferred-size arrays. It also works with allocatable components, e.g
call two(t%a)
where "a" is "allocatable, dimension(:)".
--
Summary: Unnecessary copying of array descriptors
Product: gcc
Version: 4.5.0
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=41911
