[committed] Fix OpenMP linear (ref ()) handling in Fortran FE (PR fortran/84418)
Jakub Jelinek
jakub@redhat.com
Fri Feb 16 22:45:00 GMT 2018
Hi!
For ref linear modifier, we don't want to set OMP_CLAUSE_LINEAR_STEP
to the last_step constant converted to the scalar type, but to
a sizetype last_step times the size of the referenced type.
I'm trying to resolve on omp-lang what to do about arrays with descriptors
or scalar allocatables/pointers.
Bootstrapped/regtested on x86_64-linux and i686-linux, committed to trunk,
queued for backports to 7.4.
2018-02-16 Jakub Jelinek <jakub@redhat.com>
PR fortran/84418
* trans-openmp.c (gfc_trans_omp_clauses): For OMP_CLAUSE_LINEAR_REF
kind set OMP_CLAUSE_LINEAR_STEP to TYPE_SIZE_UNIT times last_step.
* libgomp.fortran/pr84418-1.f90: New test.
* libgomp.fortran/pr84418-2.f90: New test.
--- gcc/fortran/trans-openmp.c.jj 2018-01-03 10:20:22.811538381 +0100
+++ gcc/fortran/trans-openmp.c 2018-02-16 15:49:02.500756068 +0100
@@ -1949,9 +1949,32 @@ gfc_trans_omp_clauses (stmtblock_t *bloc
}
else
{
- tree type = gfc_typenode_for_spec (&n->sym->ts);
- OMP_CLAUSE_LINEAR_STEP (node)
- = fold_convert (type, last_step);
+ if (kind == OMP_CLAUSE_LINEAR_REF)
+ {
+ tree type;
+ if (n->sym->attr.flavor == FL_PROCEDURE)
+ {
+ type = gfc_get_function_type (n->sym);
+ type = build_pointer_type (type);
+ }
+ else
+ type = gfc_sym_type (n->sym);
+ if (POINTER_TYPE_P (type))
+ type = TREE_TYPE (type);
+ /* Otherwise to be determined what exactly
+ should be done. */
+ tree t = fold_convert (sizetype, last_step);
+ t = size_binop (MULT_EXPR, t,
+ TYPE_SIZE_UNIT (type));
+ OMP_CLAUSE_LINEAR_STEP (node) = t;
+ }
+ else
+ {
+ tree type
+ = gfc_typenode_for_spec (&n->sym->ts);
+ OMP_CLAUSE_LINEAR_STEP (node)
+ = fold_convert (type, last_step);
+ }
}
if (n->sym->attr.dimension || n->sym->attr.allocatable)
OMP_CLAUSE_LINEAR_ARRAY (node) = 1;
--- libgomp/testsuite/libgomp.fortran/pr84418-1.f90.jj 2018-02-16 11:59:08.257377822 +0100
+++ libgomp/testsuite/libgomp.fortran/pr84418-1.f90 2018-02-16 11:56:19.761901263 +0100
@@ -0,0 +1,26 @@
+! PR fortran/84418
+! { dg-do run { target vect_simd_clones } }
+! { dg-options "-fno-inline" }
+! { dg-additional-options "-msse2" { target sse2_runtime } }
+! { dg-additional-options "-mavx" { target avx_runtime } }
+
+ real :: a(1024), b(1024), c(1024)
+ integer :: i
+ do i = 1, 1024
+ a(i) = 0.5 * i
+ b(i) = 1.5 * i
+ end do
+ !$omp simd
+ do i = 1, 1024
+ c(i) = foo (a(i), b(i))
+ end do
+ do i = 1, 1024
+ if (c(i).ne.(2 * i)) call abort
+ end do
+contains
+ real function foo (x, y)
+ real :: x, y
+ !$omp declare simd linear (ref (x, y))
+ foo = x + y
+ end function
+end
--- libgomp/testsuite/libgomp.fortran/pr84418-2.f90.jj 2018-02-16 16:02:22.858996874 +0100
+++ libgomp/testsuite/libgomp.fortran/pr84418-2.f90 2018-02-16 16:00:24.200952172 +0100
@@ -0,0 +1,35 @@
+! PR fortran/84418
+! { dg-do run { target vect_simd_clones } }
+! { dg-options "-fno-inline" }
+! { dg-additional-options "-msse2" { target sse2_runtime } }
+! { dg-additional-options "-mavx" { target avx_runtime } }
+
+ type p
+ integer :: i, j
+ end type
+ type(p) :: a(1024)
+ integer :: b(4,1024), c(1024)
+ integer :: i
+ do i = 1, 1024
+ a(i)%i = 2 * i
+ a(i)%j = 3 * i
+ b(1,i) = 4 * i
+ b(2,i) = 5 * i
+ b(3,i) = 6 * i
+ b(4,i) = 7 * i
+ end do
+ !$omp simd
+ do i = 1, 1024
+ c(i) = foo (a(i), b(:,i))
+ end do
+ do i = 1, 1024
+ if (c(i).ne.(6 * i)) call abort
+ end do
+contains
+ function foo (x, y)
+ type (p) :: x
+ integer :: y(4), foo
+ !$omp declare simd linear (ref (x, y))
+ foo = x%i + y(1)
+ end function
+end
Jakub
More information about the Fortran
mailing list