https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92056
--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Seems a bug introduced in r120581, COND_EXPR really needs to be handled like a
PHI node, which is handled as:
for (i = 0; i < gimple_phi_num_args (stmt); i++)
{
tree rhs = gimple_phi_arg (stmt, i)->def;
if (object_sizes[object_size_type][varno]
== unknown[object_size_type])
break;
if (TREE_CODE (rhs) == SSA_NAME)
reexamine |= merge_object_sizes (osi, var, rhs, 0);
else if (osi->pass == 0)
expr_object_size (osi, var, rhs);
}
and the break once we reach unknown size is what was missing in the COND_EXPR
handling code.
That said, I'm additionally wondering if determine_min_objsize shouldn't call
init_object_sizes ();
because otherwise it will handle only a very small subset of what
compute_builtin_object_size can do. get_destination_size from
gimple-ssa-sprintf.c calls it during the strlen pass if *printf is encountered,
but if it is not, then no SSA_NAME objsz computation is done.
