[Bug middle-end/96564] [11/12/13/14 Regression] New maybe use of uninitialized variable warning since r11-959

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96564

--- Comment #13 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Jeffrey A. Law from comment #12)
> So I think we could solve this with a bit of help from the alias oracle.  We
> have  the routine ptrs_compare_unequal, but points-to-null is going to get
> in the way.
> 
> I think VRP and DOM have enough information to rule out NULL for both
> objects in question.  So if we could query the points-to information,
> ignoring NULL then we could likely solve this particular bug.
> 
> Essentially VRP or DOM would prove NULL isn't in the set of possible values
> at the comparison point.  Then we query the alias information ignoring NULL.
> Voila we compute a static result for the comparison of the two pointers and
> the problematical block becomes unreachable and the bogus warning goes away.
> 
> Richi, any thoughts in viability of such an API?

We now treat pt.null conservatively and track non-null-ness derived from
range-info in it.  That means when VRP/DOM can prove a pointer is always
not NULL they can do set_ptr_nonnull (p) on it.

This means the

  /* ???  We'd like to handle ptr1 != NULL and ptr1 != ptr2
     but those require pt.null to be conservatively correct.  */

is no longer true and we could finally implement it, like with

diff --git a/gcc/tree-ssa-alias.cc b/gcc/tree-ssa-alias.cc
index e7c1c1aa624..5b6d9e0aa6a 100644
--- a/gcc/tree-ssa-alias.cc
+++ b/gcc/tree-ssa-alias.cc
@@ -479,9 +479,25 @@ ptrs_compare_unequal (tree ptr1, tree ptr2)
        }
       return !pt_solution_includes (&pi->pt, obj1);
     }
-
-  /* ???  We'd like to handle ptr1 != NULL and ptr1 != ptr2
-     but those require pt.null to be conservatively correct.  */
+  else if (TREE_CODE (ptr1) == SSA_NAME)
+    {
+      struct ptr_info_def *pi1 = SSA_NAME_PTR_INFO (ptr1);
+      if (!pi1
+         || pi1->pt.vars_contains_restrict
+         || pi1->pt.vars_contains_interposable)
+       return false;
+      if (integer_zerop (ptr2) && !pi1->pt.null)
+       return true;
+      if (TREE_CODE (ptr2) == SSA_NAME)
+       {
+         struct ptr_info_def *pi2 = SSA_NAME_PTR_INFO (ptr2);
+         if (!pi2
+             || pi2->pt.vars_contains_restrict
+             || pi2->pt.vars_contains_interposable)
+         if (!pi1->pt.null || !pi2->pt.null)
+           return !pt_solutions_intersect (&pi1->pt, &pi2->pt);
+       }
+    }

   return false;
 }


but the testcase shows the non-null-ness is only conditional which means
we'd have to use a range query above which necessarily falls back to
the global ranges given we don't have any context available here.  The
old EVRP adjusted global ranges during the walk but this is no longer done.

Note it's enough that one pointer is nonnull, so for your idea the
API could be extended with a bool one_ptr_nonnull parameter.