Hi!
As the testcase shows, we need to gimplify OMP_CLAUSE_SIZE, so that we
don't end up with SAVE_EXPR or anything similar non-gimple in it.
Bootstrapped/regtested on x86_64-linux and i686-linux, committed to trunk.
2021-06-08 Jakub Jelinek <ja...@redhat.com>
PR fortran/100965
* trans-openmp.c (gfc_omp_finish_clause): Gimplify OMP_CLAUSE_SIZE.
* gfortran.dg/gomp/pr100965.f90: New test.
--- gcc/fortran/trans-openmp.c.jj 2021-06-08 11:23:45.704331441 +0200
+++ gcc/fortran/trans-openmp.c 2021-06-08 16:28:09.433635685 +0200
@@ -1639,6 +1639,9 @@ gfc_omp_finish_clause (tree c, gimple_se
OMP_CLAUSE_SIZE (c)
= DECL_P (decl) ? DECL_SIZE_UNIT (decl)
: TYPE_SIZE_UNIT (TREE_TYPE (decl));
+ if (gimplify_expr (&OMP_CLAUSE_SIZE (c), pre_p,
+ NULL, is_gimple_val, fb_rvalue) == GS_ERROR)
+ OMP_CLAUSE_SIZE (c) = size_int (0);
if (c2)
{
OMP_CLAUSE_CHAIN (c2) = OMP_CLAUSE_CHAIN (last);
--- gcc/testsuite/gfortran.dg/gomp/pr100965.f90.jj 2021-06-08
16:33:29.701355827 +0200
+++ gcc/testsuite/gfortran.dg/gomp/pr100965.f90 2021-06-08 16:33:07.876647478
+0200
@@ -0,0 +1,16 @@
+! PR fortran/100965
+! { dg-do compile }
+
+implicit none
+ character(len=:), allocatable :: s
+ logical :: l
+ !$omp target map(from: l)
+ l = allocated (s)
+ !$omp end target
+ if (l) stop 1
+
+ !$omp target map(from: l)
+ l = allocated (s)
+ !$omp end target
+ if (l) stop 2
+end
Jakub
