https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121933
--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
So looking into tree-nested.cc I see this:
/* The original user label may also be use for a normal goto, therefore
we must create a new label that will actually receive the abnormal
control transfer. This new label will be marked LABEL_NONLOCAL; this
mark will trigger proper behavior in the cfg, as well as cause the
(hairy target-specific) non-local goto receiver code to be generated
when we expand rtl. Enter this association into var_map so that we
can insert the new label into the IL during a second pass. */
So maybe during gimplification we should clear LABEL_NONLOCAL.
Or better yet don't set DECL_NONLOCAL in the C front-end.
(In reply to Andrew Pinski from comment #4)
> (In reply to Andrew Pinski from comment #3)
> > I am suspecting r16-3747-gafa74d37e8170d exposed this.
So build_external_ref was not called on a label decl from a goto statement:
tree name = c_parser_peek_token (parser)->value;
if (parser->omp_metadirective_state)
name = mangle_metadirective_region_label (parser, name);
stmt = c_finish_goto_label (loc, name);
c_parser_consume_token (parser);
Likewise for &&:
```
/* Refer to the address of a label as a pointer. */
c_parser_consume_token (parser);
if (c_parser_next_token_is (parser, CPP_NAME))
{
ret.value = finish_label_address_expr
(c_parser_peek_token (parser)->value, op_loc);
set_c_expr_source_range (&ret, op_loc,
c_parser_peek_token (parser)->get_finish
());
c_parser_consume_token (parser);
}
else
{
c_parser_error (parser, "expected identifier");
ret.set_error ();
}
```
So I think the fix is in mark_decl_used:
```
if (nonloc_p)
DECL_NONLOCAL (ref) = 1;
```
Should be:
```
/* Label decls are handled by the un-nesting code in tree-nested.cc. */
if (nonloc_p && TREE_CODE (ref) != LABEL_DECL)
DECL_NONLOCAL (ref) = 1;
```
