[Bug middle-end/87296] [8/9 Regression] -Wstringop-overflow false positive due to bogus MEM_REF type

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87296

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
The MEM_REF type is significant only when actually accessing that memory, for
&MEM_REF[...]; it is insignificant and you really shouldn't assume anything
from it except for the address computation (like if it was _7 + 4).
Not really sure why we don't just fold &MEM_REF[x, y] to x p+ y though if x is
SSA_NAME.

Another example:

struct S { char a[4]; char b[6]; char c[2]; };
struct T { char a[10]; char b[4]; char c[2]; };
union U { struct S s; struct T t; };

void bar (union U *);
void baz (char *, char *, char *);

void
foo (union U *u)
{
  char (*q)[4] = &u->s.a;
  q++;
  baz ((char *) q, &u->s.b[2], &u->s.b[6]);
  bar (u);
  __builtin_strncpy (&u->t.a[4], "abcdef", 6);
}

Because pointer conversions are useless, one pointer can be replaced for
another one which has provenly the same value, but might point to different
fields etc.