Hi!
I've committed following fix for two issues revealed e.g. by valgrind
on some of the udr*.f90 testcases.
buffer could be uninitialized, and gfc_free_omp_udr could free symbols
in the combiner_ns (or initializer_ns) when freeing those whole namespaces,
so if some symbols were queued for commit/undo, it is bad to first free
the symbols and then try to undo them.
Bootstrapped/regtested on x86_64-linux and i686-linux, committed to trunk
and 4.9 branch.
2014-08-14 Jakub Jelinek <ja...@redhat.com>
PR fortran/62076
* openmp.c (gfc_match_omp_clauses): When failed to match
operator name, defined op name or name, set buffer to
empty string. Don't call gfc_find_omp_udr if buffer is empty
string.
(gfc_match_omp_declare_reduction): Call gfc_undo_symbols ()
before calling gfc_free_omp_udr.
--- gcc/fortran/openmp.c.jj 2014-08-12 15:42:44.239327384 +0200
+++ gcc/fortran/openmp.c 2014-08-14 12:17:20.554565218 +0200
@@ -464,7 +464,11 @@ gfc_match_omp_clauses (gfc_omp_clauses *
|| !gfc_add_intrinsic (&sym->attr, NULL)))
rop = OMP_REDUCTION_NONE;
}
- gfc_omp_udr *udr = gfc_find_omp_udr (gfc_current_ns, buffer, NULL);
+ else
+ buffer[0] = '\0';
+ gfc_omp_udr *udr
+ = (buffer[0]
+ ? gfc_find_omp_udr (gfc_current_ns, buffer, NULL) : NULL);
gfc_omp_namelist **head = NULL;
if (rop == OMP_REDUCTION_NONE && udr)
rop = OMP_REDUCTION_USER;
@@ -1240,6 +1244,7 @@ gfc_match_omp_declare_reduction (void)
syntax:
gfc_current_locus = old_loc;
gfc_current_ns = combiner_ns->parent;
+ gfc_undo_symbols ();
gfc_free_omp_udr (omp_udr);
return MATCH_ERROR;
}
Jakub
