C FE question: Is this test even valid?

[Diego Novillo]

With the memory SSA patches we fail gcc.dg/noncompile/920507-1.c:

int *
x(void)
{
 register int *a asm("unknown_register");  /* { dg-error "invalid
register" } */
 int *v[1] = {a};
 return v[1];
}

The expected error message on the invalid register used for 'a' does not
trigger because the store into v[0] is dead and the the check for
register validity is done during instantiation of RTL decl registers (in
the case of dead stores, no RTL is instantiated for 'a' because it
completely disappears from the IL stream).

This happens simply because we are never referencing v[0]:

x ()
{
  int * v[1];
  register int * a __asm__ (*unknown_register);
  int * D.1526;
  int * a.0;

  # VUSE <.MEM_2(D)>
  a.0_1 = a;

  # SFT.1_3 = VDEF <.MEM_2(D)>
  v[0] = a.0_1;

  # VUSE <.MEM_2(D)>
  D.1526_4 = v[1];

  return D.1526_4;
}

Notice how the load from v[1] is (naturally) not affected by the store
to v[0].

Seems to me that this test should be changed to 'return v[0]'.  In that
case, the error does trigger.

One could also argue that this check should be done closer to the FE,
but that's orthogonal to the validity of this test case.

Thanks.