[Bug c/64843] miscompilation of atomic_fetch_add on atomic pointer type

["joseph at codesourcery dot com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64843

--- Comment #3 from joseph at codesourcery dot com <joseph at codesourcery dot com> ---
The first question is whether this code is actually valid.  C11 says "All 
of these operations are applicable to an object of any atomic integer 
type.", not mentioning pointer types as valid, but then refers to address 
types.

Then, if it's valid to use pointer types here, something like the 
following (untested, and all four of the _add and _sub macros would need 
similar changes) should work without changing the built-in function 
semantics:

/* EXPR1 if it has a pointer type, otherwise EXPR2.  */
#define __atomic_ptr_choose(EXPR1, EXPR2) \
  __builtin_choose_expr (__builtin_classify_type (EXPR1) == 5, \
             (EXPR1), (EXPR2))

/* The size of *EXPR if EXPR has a pointer type, 1 otherwise.  */
#define __atomic_ptr_size(EXPR) \
  ((__PTRDIFF_TYPE__) \
   sizeof (*(__typeof (__atomic_ptr_choose (EXPR, (char *) 0))) 0))

#define atomic_fetch_add(PTR, VAL) \
  __extension__ \
  ({ \
    __auto_type __atomic_fetch_add_ptr = (PTR); \
    __atomic_fetch_add (__atomic_fetch_add_ptr, \
            (VAL) * __atomic_ptr_size (*__atomic_fetch_add_ptr), \
            __ATOMIC_SEQ_CST); \
  })