[Bug tree-optimization/67671] New: restrict pointer reference looses restrict

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67671

            Bug ID: 67671
           Summary: restrict pointer reference looses restrict
           Product: gcc
           Version: 6.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: vries at gcc dot gnu.org
  Target Milestone: ---

Consider test-case test.c:
...
void
f (int *__restrict__ &__restrict__ p)
{
  *p = 1;
}
...

When compiling the test-case with g++ -O2, we find one clique/base annotation
at ealias:
...
void f(int* __restrict__&) (intD.9 * restrict & restrict pD.2252)
{
  intD.9 * _3;

  # VUSE <.MEM_1(D)>
  # PT = nonlocal escaped
  _3 = MEM[(intD.9 * restrict &)p_2(D) clique 1 base 1];

  # .MEM_4 = VDEF <.MEM_1(D)>
  *_3 = 1;
...

If we write an equivalent example using a struct with pointer field instead:
...
struct ps {
  int *__restrict__ p;
}

f (struct ps &__restrict__ ps)
{
  *(ps.p) = 1;
}
...
we get two clique/base annotations at ealias (after fixing PR67666).

I think the two examples are similar enough to expect two clique/base
annotations at ealias for the first example.

Using this demonstrator patch, we manage that:
...
diff --git a/gcc/tree-ssa-structalias.c b/gcc/tree-ssa-structalias.c
index 94016b9..fa7108a 100644
--- a/gcc/tree-ssa-structalias.c
+++ b/gcc/tree-ssa-structalias.c
@@ -5684,6 +5684,9 @@ create_variable_info_for_1 (tree decl, const char *name)
       vi->fullsize = tree_to_uhwi (declsize);
       vi->size = vi->fullsize;
       vi->is_full_var = true;
+      if (POINTER_TYPE_P (TREE_TYPE (decl))
+         && TYPE_RESTRICT (TREE_TYPE (decl)))
+       vi->only_restrict_pointers = 1;
       fieldstack.release ();
       return vi;
     }
...