[Bug c++/67376] [5/6 regression] Comparison with pointer to past-the-end of array fails inside constant expression

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67376

--- Comment #10 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
As for the non-NULL checks,
struct A { int e[2]; };
constexpr A a { { 0, 1 } };
static_assert (a.e != 0, "");
static_assert (&a.e[1] != 0, "");
static_assert (&a.e[2] != 0, "");
static_assert (a.e + 1 != 0, "");
static_assert (a.e + 2 != 0, "");
static_assert (&a.e[-1] != 0, "");
static_assert (&a.e[3] != 0, "");
the first 3 static_asserts are handled by match.pd:
 (simplify
  (cmp (convert? addr@0) integer_zerop)
  (if (tree_single_nonzero_warnv_p (@0, NULL))
   { constant_boolean_node (cmp == NE_EXPR, type); })))
case.  The last two as well, but those are invalid and we really should be
diagnosing that in the C++ FE somewhere.
The a.e + 1 != 0 and a.e + 2 != 0 cases we used to fold them because
try_move_mult_to_index would fold them into ADDR_EXPR of some field, which
intentionally no longer happens.  So, I guess we want instead some
simplification of POINTER_PLUS_EXPR<ADDR_EXPR<something>, INTEGER_CST> that
would say use get_inner_reference on something, add (in offset_int?)  the
constant offset to it and determine if it is always non-zero.

The question is what rule should we use.  Any positive constant offset from a
known non-zero base address?  Something different (like only positive constant
offset from 0 up to and including the size of the var)?  Perhaps if the
constant offset modulo alignment of the var is known to be non-zero always, no
matter if the base var has known non-zero base address or not?  Or just
whenever the constant offset is > 0 (or optionally also smaller than size of
the var)?  I mean, for &a.e + 1, even if &a is NULL, &a.e + 1 still would be
non-NULL, unless the a var is partly at the end of the address space and partly
at the beginning.