Document #: | P3293R1 [Latest] [Status] |
Date: | 2024-10-13 |
Project: | Programming Language C++ |
Audience: |
EWG |
Reply-to: |
Peter Dimov <pdimov@gmail.com> Dan Katz <dkatz85@bloomberg.net> Barry Revzin <barry.revzin@gmail.com> Daveed Vandevoorde <daveed@edg.com> |
Since [P3293R0], noted that &[:base:]
cannot work for virtual base classes. Talking about arrays. Added
wording.
There are many contexts in which it is useful to perform the same operation on each subobject of an object in sequence. These include serialization or formatting or hashing.
[P2996R6] seems like it gives us an ideal solution to this problem, in the form of being able to iterate over all the subobjects of an object and splicing accesses to them. However, it is not quite complete:
template <class T, class F> void for_each_subobject(T const& obj, F f) { template for (constexpr auto sub : subobjects_of(^T)) { (obj.[:sub:]); // this is valid syntax for non-static data members f// but is invalid for base classes subobjects } }
Instead we have to handle bases separately from the non-static data members:
template <class T, class F> void for_each_subobject(T const& obj, F f) { template for (constexpr auto base : bases_of(^T)) { (static_cast<type_of(base) const&>(obj)); f} template for (constexpr auto sub : nonstatic_data_members_of(^T)) { (obj.[:sub:]); f} }
Except this is now a normal
static_cast
and so requires access checking, thus prohibiting accessing private base
classes.
We could avoid access checking by using a C-style cast:
template <class T, class F> void for_each_subobject(T const& obj, F f) { template for (constexpr auto base : bases_of(^T)) { ((typename [: type_of(base) :]&)obj); f} template for (constexpr auto sub : nonstatic_data_members_of(^T)) { (obj.[:sub:]); f} }
But this opens up other problems: I forgot to write
const
and so
now I accidentally cast away
const
-ness
unintentionally. Oops. Not to mention that this cast actually works
regardless of whether base
refers to
a base class of T
, so it’s not
exactly the best programming practice.
On top of that, both the
static_cast
and C-style cast approaches suffer from having to correctly spell the
destination type - which requires manually propagating the const-ness
and value category of the object.
The way to avoid all of these problems is to just defer to a function template:
template <std::meta::info M, class T> constexpr auto subobject_cast(T&& arg) -> auto&& { constexpr auto stripped = remove_cvref(^T); if constexpr (is_base(M)) { static_assert(is_base_of(type_of(M), stripped)); return (typename [: copy_cvref(^T, type_of(M)) :]&)arg; } else { static_assert(parent_of(M) == stripped); return ((T&&)arg).[:M:]; } } template <class T, class F> void for_each_subobject(T const& obj, F f) { template for (constexpr auto sub : subobjects_of(^T)) { (subobject_cast<sub>(obj)); f} }
But this feels a bit silly? Why should we have to write this?
There are three kinds of subobjects as specified in 6.7.2 [intro.object]:
2 Objects can contain other objects, called subobjects. A subobject can be a member subobject ([class.mem]), a base class subobject ([class.derived]), or an array element.
Currently, the reflection proposal only supports splicing member subobjects. Let’s go over the other two.
Unlike member subobjects, there is no way today to access a
base class subobject directly outside of one of the casts described
above. Part of the reason for this is that while a data member is always
just an identifier
, a base
class subobject can have an arbitrary complex name. Reflection allows us
to sidestep the question of complex names since we just have a
reflection to the appropriate base class subobject already, so it
becomes a matter of asking what obj.[:base:]
means.
But what else could it mean? We argue that the obvious, useful, and
only possible meaning of this syntax would be to access the appropriate
base class subobject (in the same way that obj.[:nsdm:]
— where nsdm
represents a non-static
data member — is access to that data member).
Additionally, &[:base:]
where base
represents a base class
B
of type
T
could yield a
B T::*
with
appropriate offset. Likewise, there is no way to directly get such a
member pointer today. But that would be a useful bit of functionality to
add. That is, unless base
represents
a reflection of a virtual base class subobject, which wouldn’t be
representable as a pointer to member.
The only reason [P2996R6] doesn’t support splicing base class subobjects is the lack of equivalent language support today. This means that adding this support in reflection would mean that splicing can achieve something the language cannot do natively. But we don’t really see that as a problem. Reflection is already allowing all sorts of things that the language cannot do natively. What’s one more?
The more complicated question is array elements. subobjects_of(^int[4])
needs to return four subobjects, which would effectively each represent
“index i
into int[4]
”
for each i
in [0, 1, 2, 3]
.
We would want both the index and the type here. That, in of itself,
isn’t any different from dealing with the non-static data members of a
class. These subobjects likewise have offsets, types, etc, in the same
way.
The challenge is more in terms of access. With base class subobjects,
we talked about how there is no non-splice equivalent to obj.[:base:]
.
But today there isn’t any even any such valid syntax for arrays today at
all! Indeed, the standard refers to the expression
E1.E2
as
class member access (7.6.1.5 [expr.ref]), and arrays are
not actually classes. That adds complexity.
Moreover, arrays have the additional problem that the number of array
elements can explode quite quickly. int[1'000]
already has 1000 subobjects. That’s expensive. Especially since array
elements aren’t quite the same as the other kinds of subobjects — we
know they are all the same type. All the use-cases that we have, unlike
for base class subobjects, would want to treat arrays separately
anyway.
That is, being able to splice base class subobjects is small language extension with good bang for the buck. Being able to splice array subobjects requires being able to even represent array subobjects as reflections (which currently does not exist) as well as being able to extend class member access to support arrays. Not to mention pointers to members? It’s a much larger change with a much smaller benefit.
We propose two language changes:
obj.[:base:]
(where base
is a reflection of a
base class of the type of obj
) as
being an access to that base class subobject, in the same way that obj.[:nsdm:]
(where nsdm
is a reflection of a
non-static data member) is an access to that data member.&[:base:]
where base
is a reflection of a base
class B
of type
T
should yield a
B T::*
with
appropriate offset. Unless base
is a
reflection of a virtual base class, which wouldn’t really be
representable as a pointer to member.We argue that these are the obvious, useful, and only possible meanings of these syntaxes, so we should simply support them in the language.
The only reason this isn’t initially part of [P2996R6] is that while there is
a way to access a data member of an object directly (just
obj.mem
),
there is no way to access a base class subobject directly
outside of one of the casts described above.
We then additionally propose to back subobjects_of()
(and get_public_subobjects()
)
that [P2996R6] removed. These were removed
because iterating over all the subobjects uniformly wasn’t
really possible until these language additions.
Handle base class splices in 7.6.1.5 [expr.ref]/7-8:
7 Otherwise, one of the following rules applies.
(7.4b) If E2 designates a direct, non-virtual base class subobject
B
and type ofB
isT
, the expression designates the corresponding base class subobject of the object designated by the first expression. If E1 is an lvalue, then E1.E2 is an lvalue; otherwise E1.E2 is an xvalue. The type of E1.E2 is “cv T
”. [ Note 1: This can only occur in an expression of the forme1.[:e2:]
wheree2
is a reflection designating a base class subobject. — end note ][ Example 1:— end example ]struct B { int b; }; struct D : B { int d; }; constexpr int f() { = {1, 2}; D d & b = d.[: bases_of(^D)[0] :]; B.b += 10; b.[: ^D::d :] += 1; dreturn d.b * d.d; } static_assert(f() == 33);
- (7.5) If E2 is a member enumerator and the type of E2 is T, the expression E1.E2 is a prvalue of type T whose value is the value of the enumerator.
8 If
E2
designates a non-static memberM
, the program is ill-formed if the class of whichM
is directly a member is an ambiguous base ([class.member.lookup]) of the naming class ([class.access.base]) ofM
. IfE2
designates a base class subobjectB
, the program is ill-formed ifB
is not a direct base class of the naming class ofB
.
Handle base class pointers to members in 7.6.2.2 [expr.unary.op]:
3 The operand of the unary
&
operator shall be an lvalue of some typeT
.
- (3.1) If the operand is a
qualified-id
orsplice-expression
designating a non-static or variant member of some classC
, other than an explicit object member function, the result has type “pointer to member of classC
of typeT
” and designatesC::m
.
- (3.1+) If the operand is a
splice-expression
designated a base class subobject of some classC
of typeT
, other than a virtual base class subobject, the result has type pointer to member of classC
of typeT
and designates that base class subobject.
- (3.2) Otherwise, the result has type “pointer to
T
” and points to the designated object (6.7.1 [intro.memory]) or function (6.8.4 [basic.compound]). If the operand designates an explicit object member function (9.3.4.6 [dcl.fct]), the operand shall be aqualified-id
or asplice-expression
.[ Example 2:— end example ]struct B { int b; }; struct D : B { int d; }; constexpr D d = {1, 2}; constexpr int B::*pb = &[: bases_of(^D)[0] :]; constexpr int D::*pd = &D::d; static_assert(&(d.*pb) == &static_cast<B&>(d)); static_assert(d.*pd == 2);
4 A pointer to member is only formed when an explicit
&
is used and its operand is aqualified-id
orsplice-expression
not enclosed in parentheses.
Add to meta.synop:
namespace std::meta { // ... // [meta.reflection.member.queries], reflection member queries consteval vector<info> members_of(info r); consteval vector<info> bases_of(info type); consteval vector<info> static_data_members_of(info type); consteval vector<info> nonstatic_data_members_of(info type);+ consteval vector<info> subobjects_of(info type); consteval vector<info> enumerators_of(info type_enum); consteval vector<info> get_public_members(info type); consteval vector<info> get_public_static_data_members(info type); consteval vector<info> get_public_nonstatic_data_members(info type); consteval vector<info> get_public_bases(info type);+ consteval vector<info> get_public_subobjects(info type); // ... }
Add to [meta.reflection.member.queries] in the appropriate spot:
consteval vector<info> nonstatic_data_members_of(info type);
12 Constant When:
dealias(type)
represents a complete class type.13 Effects: If
dealias(type)
represents a class template specialization with a reachable definition, the specialization is instantiated.14 Returns: A
vector
containing the reflections of the direct non-static data members of the type represented bydealias(type)
, in the order in which they are declared.consteval vector<info> subobjects_of(info type);
14+1 Constant When:
dealias(type)
represents a complete class type.14+2 Effects: If
dealias(type)
represents a class template specialization with a reachable definition, the specialization is instantiated.14+3 Returns: A
vector
containing all the reflections inbases_of(type)
followed by all the reflections innonstatic_data_members_of(type)
.consteval vector<info> get_public_bases(info type);
26 Constant When:
dealias(type)
represents a complete class type.27 Effects: If
dealias(type)
represents a class template specialization with a reachable definition, the specialization is instantiated.28 Returns: A
vector
containing each element,e
, ofbases_of(type)
such thatis_public(e)
istrue
, in order.consteval vector<info> get_public_subobjects_of(info type);
29 Constant When:
dealias(type)
represents a complete class type.30 Effects: If
dealias(type)
represents a class template specialization with a reachable definition, the specialization is instantiated.31 Returns: A
vector
containing all the reflections inget_public_bases(type)
followed by all the reflections inget_public_nonstatic_data_members(type)
.