[Bug rtl-optimization/69887] [4.9/5/6 Regression] gcc ICE at -O1 and above on x86_64-linux-gnu in mark_jump_label_1

["jakub at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69887

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ebotcazou at gcc dot gnu.org,
                   |                            |jakub at gcc dot gnu.org

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
void *a[6];

void
foo ()
{
  __builtin_setjmp (a);
  __builtin_longjmp (a, 1);
}

The problem is that pass_expand::execute in
          /* At the moment not all abnormal edges match the RTL
             representation.  It is safe to remove them here as
             find_many_sub_basic_blocks will rediscover them.
             In the future we should get this fixed properly.  */
          if ((e->flags & EDGE_ABNORMAL)
              && !(e->flags & EDGE_SIBCALL))
            remove_edge (e);
          else
            ei_next (&ei);
removes the abnormal edges, but find_many_sub_basic_blocks doesn't actually
recreate any corresponding edge from the REG_NON_LOCAL_GOTO JUMP_INSN (the
label is not marked as forced_label, just LABEL_PRESERVED_P), thus the cfg
cleanup at the end of expand pass attempts to remove the
__builtin_setjmp_receiver basic block as unreachable.  As the label is
LABEL_PRESERVED_P, it doesn't actually delete the label, just turns it into
NOTE_INSN_DELETED_LABEL, and there is some insn earlier (expanded from
__builtin_setjmp_setup) that references that deleted label through LABEL_REF.
Later on CSE determines that the REG_NON_LOCAL_GOTO JUMP_INSN can only jump to
the particular (deleted) label, and bumps its LABEL_NUSES (probably rtl
checking would ICE here, non-rtl checking just silently turns that
NOTE_INSN_DELETED_LABEL into NOTE_INSN_DELETED_DEBUG_LABEL.
Thus, a quick hack could be to tweak cse.c to follow mark_jump_label_1 and
don't touch LABEL_NUSES on deleted labels.  But, it still looks wrong to me,
we'd then attempt to jump into a middle of basic block.
So I bet we want to do something during find_many_sub_basic_blocks or elsewhere
during expansion, so that we keep the setjmp receiver basic block alive.