https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84102
Bug ID: 84102
Summary: Fails to disambiguate Fortran (non-addressable?)
global with array descriptor data
Product: gcc
Version: 8.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
Created attachment 43269
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=43269&action=edit
reduced capacita
This is split out from PR84037 with a reduced testcase. We fail to hoist the
load of D1 from
do i=1,Ng1 ! .. and multiply charge with x
do j=1,Ng2
X(i,j) = X(i,j) * D1 * (i-(Ng1+1)/2.0_dp)
end do
end do
which ultimatively causes a runtime alias check for vectorization to
disambiguate
D1 and X.
Note X is allocate()d and properly gets a malloc DECL from PTA. But the
function calls solve(X,Y0) and GCC believes this call may alter X.data
(and other fields of the descriptor). X is
real(kind=dp), intent(out), dimension(:,:) :: X
inside solve so that might be very well possible. But of course X may
not end up pointing to D1 for whatever details. My guess is that it's
the usual 'you can't take the address of random stuff without annotating
it in fortran'. If true this would deserve sth stronger than
!TREE_ADDRESSABLE which we only trust for TU-local variables. -fwhole-program
helps here to properly constrain D1.