[Bug tree-optimization/81008] New: missing optimization for restricted pointers passed to functions

[msebor at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81008

            Bug ID: 81008
           Summary: missing optimization for restricted pointers passed to
                    functions
           Product: gcc
           Version: 7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: msebor at gcc dot gnu.org
  Target Milestone: ---

While exploring ways to annotate C++ classes or member functions to indicate
that accessing an object of the class does not change a different object of
that class I noticed that GCC doesn't take full advantage of the basic
guarantee on restrict-qualified pointers: that if one is used to access an
object, the same object may only be modified through an [lvalue] based on the
same pointer.

In the example below, the call to foo() is guaranteed not to modify the object
at *p because p is a restricted pointer and there is no other pointer based on
p can exist.

Clang emits optimal code for the test case (i.e., just a jump to foo()).

$ cat t.c && gcc -O3 -S -Wall -Wrestrict -fdump-tree-optimized=/dev/stdout t.c

void f (void);

void g (int* restrict p)
{
  int x = *p;
  f ();
  if (x != *p)
    __builtin_abort ();
}

;; Function g (g, funcdef_no=0, decl_uid=1795, cgraph_uid=0, symbol_order=0)

g (int * restrict p)
{
  int x;
  int _1;

  <bb 2> [100.00%]:
  x_4 = *p_3(D);
  f ();
  _1 = *p_3(D);
  if (_1 != x_4)
    goto <bb 3>; [0.04%]
  else
    goto <bb 4>; [99.96%]

  <bb 3> [0.04%]:
  __builtin_abort ();

  <bb 4> [99.96%]:
  return;

}


In the above, the only way f() could modify *p is like so:

  int *q;
  void f (void) { *q = 0; }

and

  void h (void)
  {
    int d = 0;
    q = &d;
    g (&d);
  }

but such a modification is undefined thanks to the access to d in f() through a
restricted pointer.