[Bug ipa/87843] [9 Regression] SPEC miscompilation of 403.gcc and 502.gcc_r benchmarks

[hubicka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87843

--- Comment #5 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
I think this is caused by misoptimizing
     void **x;
     void *info ATTRIBUTE_UNUSED;
{
  cselib_val *v = (cselib_val *)*x;
  struct elt_loc_list **p = &v->locs;
  int had_locs = v->locs != 0;

  while (*p)
    {
      if (references_value_p ((*p)->loc, 1))
        unchain_one_elt_loc_list (p);
      else
        p = &(*p)->next;
    }

  if (had_locs && v->locs == 0)
    {
      n_useless_values++;
      values_became_useless = 1;
    }
  return 1;
}

where fre1 after type cleaning concludes that v->locs is unchanged while it is
unchanged in:

static void
unchain_one_elt_loc_list (pl)
     struct elt_loc_list **pl;
{
  struct elt_loc_list *l = *pl;

  *pl = l->next;
  l->next = empty_elt_loc_lists;
  empty_elt_loc_lists = l;
}

So it seems that in some cases alias analysis gets lost in pointer dereferences
while we are still in early optimization. 

Perhaps we want to produce alias sets earlier?

Honza