The testcase for PR93423 did a double free, which caused
an ICE. That's reported in PR96041.
Slightly frustrated by the FAIL in the testsuite,
I decided to debug and, hopefully, fix this.
The problem is related to putting the symtree
into a sub namespace of the symbol's ns. That's fixed up
by copying things around – except in the error case where
all those fixups are undone. Thus, when the symbol tree
is deleted, the parent's sym->formal->sym is also deleted,
causing an ICE in resolve_formal_arguments.
Hopefully, I got this all right...
I see still one memory leak for a symbol in module.c
according to valgrind, but I don't know whether it is
related to those symbols. (There are a lot of other leaks,
mostly related to polymorphism (vtab etc.).)
OK for the trunk?
Tobias
Fortran: Avoid double-free with parse error (PR96041, PR93423)
gcc/fortran/
PR fortran/96041
PR fortran/93423
* decl.c (gfc_match_submod_proc): Avoid later double-free
in the error case.
gcc/fortran/decl.c | 9 +++++++++
1 file changed, 9 insertions(+)
diff --git a/gcc/fortran/decl.c b/gcc/fortran/decl.c
index c612b492f3e..326e6f5db7a 100644
--- a/gcc/fortran/decl.c
+++ b/gcc/fortran/decl.c
@@ -9819,6 +9819,15 @@ gfc_match_submod_proc (void)
if (gfc_match_eos () != MATCH_YES)
{
+ /* Unset st->n.sym. Note: in reject_statement (), the symbol changes are
+ undone, such that the st->n.sym->formal points to the original symbol;
+ if now this namespace is finalized, the formal namespace is freed,
+ but it might be still needed in the parent namespace. */
+ gfc_symtree *st = gfc_find_symtree (gfc_current_ns->sym_root, sym->name);
+ st->n.sym = NULL;
+ gfc_free_symbol (sym->tlink);
+ sym->tlink = NULL;
+ sym->refs--;
gfc_syntax_error (ST_MODULE_PROC);
return MATCH_ERROR;
}
