[Bug rtl-optimization/103860] [9/10/11/12 Regression] wrong code at -O3 with -fPIC on x86_64-linux-gnu

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103860

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
In the:
  while (!vec.is_empty () && pro != entry)
    {
      while (pro != entry && !can_get_prologue (pro, prologue_clobbered))
        {
          pro = get_immediate_dominator (CDI_DOMINATORS, pro);

          if (bitmap_set_bit (bb_with, pro->index))
            vec.quick_push (pro);
        }

      basic_block bb = vec.pop ();
      if (!can_dup_for_shrink_wrapping (bb, pro, max_grow_size))
        while (!dominated_by_p (CDI_DOMINATORS, bb, pro))
          {
            gcc_assert (pro != entry);

            pro = get_immediate_dominator (CDI_DOMINATORS, pro);

            if (bitmap_set_bit (bb_with, pro->index))
              vec.quick_push (pro);
          }

      FOR_EACH_EDGE (e, ei, bb->succs)
        if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
            && bitmap_set_bit (bb_with, e->dest->index))
          vec.quick_push (e->dest);
    }
loop, initially we try as pro bb 7 for which can_get_prologue (pro,
prologue_clobbered) is true.
We iterate for a while, until bb_with is 3, 4, 5, 6, 7, 8, pro still 7,
and we pop bb 6, for which can_dup_for_shrink_wrapping is false.
vec at this point is empty.  As bb 6 isn't dominated by pro 7, we change
pro to the immediate dominator of bb 7 which is bb 5, and as it already is in
bb_with bitmap, we don't push anything.  bb 6 also has no successors (it traps
at the end), so we don't push there anything either.  And because vec.is_empty
() is true, the outer loop doesn't iterate any longer and so nothing checks if
can_get_prologue (pro, prologue_clobbered) (which is false for bb 5).