https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88983
--- Comment #3 from Marek Polacek <mpolacek at gcc dot gnu.org> ---
I think I see the problem: we're evaluating the body of the switch, and cond is
"1" so we're jumping over everything until we find "case 1":
if (1)
{
case 1:;
return <retval> = 1;
}
else
{
default:;
}
4681 case COND_EXPR:
4682 if (jump_target && *jump_target)
here jump target is "1"
4683 {
4684 /* When jumping to a label, the label might be either in the
4685 then or else blocks, so process then block first in skipping
4686 mode first, and if we are still in the skipping mode at its
end,
4687 process the else block too. */
4688 r = cxx_eval_constant_expression (ctx, TREE_OPERAND (t, 1),
4689 lval, non_constant_p,
overflow_p,
4690 jump_target);
we found "case 1" in the then branch, but the next statement was "return", so
we have a new jump_target, which...
4691 if (*jump_target)
...confuses this condition, and we go looking to the else branch...
4692 r = cxx_eval_constant_expression (ctx, TREE_OPERAND (t, 2),
4693 lval, non_constant_p,
overflow_p,
4694 jump_target);
...where we crash.
