The C++ FE drops qualifiers on pointer dereference

[Richard Guenther <rguenther@suse.de>]

The C++ frontend drops qualifiers when performing pointer dereference:

void foo(int * __restrict p, int * __restrict__ q, int i)
{
  *q = p[i];
}

;; Function void foo(int*, int*, int) (null)
;; enabled by -tree-original


<<cleanup_point <<< Unknown tree: expr_stmt
  (void) (*(int *) q = *((int *) p + (long unsigned int) ((long unsigned 
int) i * 4))) >>>
>>;

while this isn't a big issue for the plain *q (the cast is stripped
as useless during gimplification), the second cast results in the
gimplifier generating an unqualified temporary to do the dereference:

void foo(int*, int*, int) (int * restrict p, int * restrict q, int i)
{
  long unsigned int D.2088;
  long unsigned int D.2089;
  int * D.2090;
  int D.2091;

  D.2088 = (long unsigned int) i;
  D.2089 = D.2088 * 4;
  D.2090 = p + D.2089;
  D.2091 = *D.2090;
  *q = D.2091;
}

The C frontend seems to get away without doing this conversion, the
C++ frontend does it here:

typeck.c:decay_conversion

  /* [basic.lval]

     Non-class rvalues always have cv-unqualified types.  */
  type = TREE_TYPE (exp);
  if (!CLASS_TYPE_P (type) && cp_type_quals (type))
    exp = build_nop (TYPE_MAIN_VARIANT (type), exp);

called from

#1  0x0000000000657c35 in cp_build_indirect_ref (ptr=0x7ffff7edcf30, 
    errorstring=0x120f491 "unary *", complain=3)
    at /space/rguenther/src/svn/trunk/gcc/cp/typeck.c:2493
2490      if (ptr == current_class_ptr)
2491        return current_class_ref;
2492
2493      pointer = (TREE_CODE (TREE_TYPE (ptr)) == REFERENCE_TYPE
2494                 ? ptr : decay_conversion (ptr));
2495      type = TREE_TYPE (pointer);

The C++ frontend in this way currently pessimizes code by emitting
dereferences of non-restrict qualified pointers where the user
used a restrict qualified pointer.

Where can this be fixed?  I currently have a hack to not strip
restrict qualifiers in decay_conversion, but that feels a little
too hackish.  What's the important piece of decay_conversion
that is necessary for indirect refs?

Thanks,
Richard.