Joseph Thomson <[email protected]>
16 April 2019
In this document, I outline why the the C++ Core Guidelines' current recommendations regarding the use of pointers in modern C++ code are not ideal. I go on to argue for the addition of two types, retained<T>, and optional_ref<T>, to the Guideline Support Library. Finally, I discuss the role of owner<T> and not_null<T> in the validation of code by static analysis tools, and suggest an alternative approach to pointer annotation.
Working implementations of the proposed classes can be found here, and a full test suite can be found here.
This proposal does not recommend the replacement of all uses of pointers in modern C++. Low-level implementations and legacy code bases will inevitably still use pointers. I instead propose that strongly typed alternatives to T* should be provided and recommended by the guidelines, especially for new high-level code. I do not recommend that static analysis tools flag all instances of T* for replacement. In fact, in the discussion of pointer annotation, I suggest that a bare T* should be understood by static analysis tools to represent a single object.
- Version 1 (9 February 2017) - Original "draft" version. Proposal was posted to the C++ Core Guidelines GitHub page, but ultimately did not gain any interest. Some fair points were made about the pragmatic approach taken to pointer annotation, but I will leave that part of the proposal as-is, as I believe it to still be good, if a little idealistic.
- Version 2 (16 April 2019) - Updated the proposal, not in hopes of pitching it again, but to redesign
observer<T>(now namedretained<T>) to maintain const-correctness, which I now believe to be the correct design after a few years of using it in practice.
Pointer types define a large range of operations. However, many of these operations have well-defined behaviour only in specific circumstances. For example (given a pointer, p, and an integral, n):
- The expression
p + nhas only has defined behaviour the resulting pointer is within the bounds of an array object (including one past the end of the array). - The expression
*phas undefined behaviour ifpis a null pointer or a past-the-end iterator. - The expression
delete phas undefined behaviour ifppoints to an object not allocated withnew. - The expression
delete[] phas undefined behaviour ifppoints to an object not allocated withnew[].
The problem is that T* is weakly typed, in that it can be used for many, unrelated purposes. Thus, use of T* violates rule I.4, which tells us to, "Make interfaces precisely and strongly typed", explaining that:
Types are the simplest and best documentation, have well-defined meaning, and are guaranteed to be checked at compile-time.
It also violates rule P.1, which advises us to, "Express ideas directly in code", noting that:
Compilers don't read comments (or design documents) and neither do many programmers (consistently). What is expressed in code has defined semantics and can (in principle) be checked by compilers and other tools.
By using T*, we fail to document our intent to the detriment of both the programmer and the compiler. T* has overly broad semantics, and our code would become both safer and clearer if instances of T* in high-level code were replaced with strongly typed alternatives. The guidelines currently recommend various such types—std::unique_ptr, std::shared_ptr, std::array, stack_array—but there are still places where the guidelines recommend the use of T*. Rule F.22 advises us to, "Use T* or owner<T*> to designate a single object", giving the reasoning:
Readability: it makes the meaning of a plain pointer clear. Enables significant tool support.
Designating T* to represent only single objects does indeed make the meaning of a T* clear and enable tool support, but only by a consensus of programmers and static analysis tools. Agreement by the compiler in the form of strong typing would greatly strengthen such meaning. This rule provides no reason why a T* in this one situation would not benefit from a strongly typed alternative. Rule Bounds.1 instructs, "Don't use pointer arithmetic. Use span instead", explaining:
Pointers should only refer to single objects, and pointer arithmetic is fragile and easy to get wrong.
spanis a bounds-checked, safe type for accessing arrays of data.
It appears that the guidelines are reasoning that, since span is provided as a substitute for pointers when used as iterators, the only use left for T* is as a reference to single objects, so by process of elimination this must be the single appropriate use for T* in modern C++ code. Again, this ignores the fact that the defined semantics of T* are too general for something that represents a "single object". For example, pointer arithmetic operations make little sense if T* is meant to point to a single object. Of course, static analysis tools could flag uses of T* that make no sense for a "single object", but this relies on the ubiquity and reliable operation of static analysis tools, the development of which has only recently begun. There is no good reason to reject what the compiler can offer us, and forego the good design practices that the guidelines themselves recommend.
If type-safety is not a compelling enough reason to provide strongly typed alternatives for T* in all situations, consider that the guidelines still recommend the use of T* as a "single object" for two conceptually distinct purposes. Moreover, the semantics of T* are not a good fit for either purpose; in fact, the ideal semantics for the two purposes are diametrically opposed. Thus, there is no way to provide a single high-level type to perform both functions.
The first use of T* is described in rule F.60, which tells us to, "Prefer T* over T& when "no argument" is a valid option", explaining:
A pointer (
T*) can be anullptrand a reference (T&) cannot, there is no valid "null reference". Sometimes havingnullptras an alternative to indicated "no object" is useful, but if it is not, a reference is notationally simpler and might yield better code.
Thus, we might implement such an "optional reference" parameter like so:
void frobnicate(widget const* ow);
Aside from the aforementioned type-safety issues, the use of T* as an optional T& is unnatural, as the semantics of pointers and references are very different. In particular, one would expect to be able to pass temporary arguments to "optional reference" parameters, but it is illegal to take the address of an rvalue:
frobnicate(&widget()); // error: cannot take address of rvalue
Ideally, one would be able to pass arguments to "optional reference" parameters using the same syntax as with T& parameters:
frobnicate(widget()); // ideal
The second use of T* is not explicitly described by the guidelines, but can be identified by considering how best to store a non-owning reference. If T& is the appropriate way to represent a non-optional reference parameter, we might consider using a T& data member to store a reference:
class foo {
public:
explicit foo(bar& b) : b(b) {}
…
private:
bar& b;
};
Unfortunately, T& data members make their containing class non-copy assignable by default. In addition, they cannot be stored in STL containers. Instead, we could use T*, but our reference is non-optional, so we had better use not_null<T*>:
class foo {
public:
explicit foo(bar& b) : b(&b) {}
…
private:
not_null<bar*> b;
};
However, there are still a number of problems with this approach. Firstly, it is not generally expected that a function taking a T& parameter will retain a pointer to its argument. The intent to retain is unclear from looking at either the function signature or the calling code:
foo f{b}; // looks like "pass by value"
Rule F.15 tells us to, "Prefer simple and conventional ways of passing information", explaining:
Using "unusual and clever" techniques causes surprises, slows understanding by other programmers, and encourages bugs.
Storing a pointer to a T& parameter could be considered "unusual" and is arguably an example of a violation of this rule, as conventionally, the semantics of T& are that of an "out" parameter, not of a retained reference. It gets worse with T const&:
class foo {
public:
explicit foo(bar const& b) : b(b) {}
…
private:
bar const& b;
};
It is even more surprising that a reference is retained, since a T const& parameter is typically a T parameter would result in an expensive copy. Furthermore, it is possible to pass temporaries to T const& parameters, which in this case will inevitably result in dangling pointers.
foo f{bar()}; // `f.b` is left dangling
We could instead take a not_null<T*> parameter, but then we would lose the compile-time enforcement of the "not null" condition (this is also a disadvantage of using not_null<T*> as a data member, though in this case the member is private, so opportunities for bugs to arise are encapsulated within the class). In addition, the calling code is indistinguishable from that of a T* "optional reference" parameter, and it is still likely to be surprising that a copy of the pointer is retained. If a function is going to store a copy of a reference or pointer parameter, the intent to do so should be made explicitly clear in both the function signature and at the call site, so that the caller is made aware that they must personally manage the lifetimes of both objects. Currently, there is no conventional way to do this in C++.
Lastly, member functions of type T*, T& or not_null<T*> break const-correctness. The data members of a read-only object are themselves read-only (unless they are marked mutable). However, the objects pointed to by read-only data members of type T* and T& are not read-only. This is the expected behaviour, but is undesirable in most high-level code. Thus, on the basis of const-correctness alone, data members of type T* and T& are unsuitable for modern, high-level C++ code.
We have identified the two uses of T* for which the guidelines do not give strongly typed alternatives:
- Retained, non-owning references
- "Optional reference" parameters
Proposed here are two class templates, retained<T> and optional_ref<T>, that can be used in place of T* in these two situations respectively.
The retained<T> class template is a pointer-like type designed as a substitute for T* wherever it is used as a non-owning "retained" of an object of type T.
class foo {
public:
explicit foo(retained<bar> b) : b(b) {}
…
private:
retained<bar> b;
};
Because retained<T> is only explicitly constructible from T&, the intention to retain a copy of the reference has to be explicitly documented at the call site, usually via the make_retained factory function:
foo f{make_retained(b)};
In addition, retained<T> disables construction from T&&, so it cannot be constructed from temporary objects.
foo f{make_retained(bar())}; // error: `make_retained(T&&)` is deleted
Furthermore, retained<T> const has read-only access to the object it references, thereby enforcing const-correctness:
void foo::frobnicate() const {
*b = bar(); // error: assignment of read-only variable
}
A side-effect of this behaviour is that retained<T> is move-only, as making it copyable would break const-correctness. Nonetheless, it is easy to make const-correct "copies" of retained<T>:
retained<bar> get() {
return make_retained(*b)
}
retained<bar const> get() const {
return make_retained(*b)
}
As a zero-overhead alternative for T*, retained<T> does not manage or track the lifetime of what it references. If automatic lifetime tracking is required, alternative approaches such as a signals and slots implementation (e.g. Boost.Signals2) or std::weak_ptr may be used.
An retained<T> has no "null" state and must therefore always point to an object. Thus, it enforces the "not null" condition at compile-time, just like T&. If the ability to represent "no object" is required, retained<T> can be combined with optional<T>:
class foo {
public:
foo() = default;
explicit foo(retained<bar> b) : b(b) {}
…
private:
optional<retained<bar>> b;
};
Or alternatively:
class foo {
public:
explicit foo(optional<retained<bar>> b = nullopt) : b(b) {}
…
private:
optional<retained<bar>> b;
};
A naive implementation of optional<retained<T>> will not have zero-overhead (most notably, it usually occupies twice as much memory as T*). However, the as-if rule should allow a zero-overhead implementation in practice if an optional<retained<T>> specialization is given "back-door" access to the pointer member internal to retained<T>, the unused null pointer state of which it can use to represent its disengaged state.
There is a soon to be standardized way to represent the concept of "optional" values in C++: std::optional<T>. Ideally, we would be able to use the optional<T&> specialization instead of T* as a way to represent "optional reference" parameters. Unfortunately, although optional<T&> was included as an auxiliary proposal to N3527, it doesn't look like it will be accepted into the standard.
The optional_ref<T> class template is an "optional" reference-like type, and is essentially a slightly modified version of optional<T&> as described in N3527.
void frobnicate_default();
void frobnicate_widget(widget const& w);
void frobnicate(optional_ref<widget const> ow) {
if (ow) {
frobnicate_widget(*ow);
} else {
frobnicate_default();
}
}
Since T& converts implicitly to optional_ref<T>, it is possible to pass temporaries to optional_ref<T const> parameters just as with T const& parameters:
frobnicate(widget());
frobnicate(nullopt);
However, the design differs somewhat from the proposed version of optional<T&>.
Allowing construction from T&& is essential to allow passing of temporary arguments to optional_ref<T> parameters. The optional proposal decided to disable construction from T&& to prevent the accidental formation of dangling references by programmers who mistakenly expect optional<T const&> to extend the lifetime of temporaries as T const& does. However, I believe this is a price worth paying to enable temporaries to be passed as optional_ref<T const> parameters, especially given that std::basic_string_view, a reference-like type with similar semantics, make the exact same compromise.
The optional<T&> auxiliary proposal discusses how the semantics for copy assignment proved controversial. They ultimately chose reference copy assignment semantics somewhat arbitrarily, copying the behaviour of std::reference_wrapper and boost::optional. The proposal mentions that most people insisted that optional<T&> be copy assignable, but no rationale is given. I speculate that this is because optional<T> is viewed by most people as inheriting the operations defined by T. Since T& is copy assignable (although copy assignment for T& has value semantics, not reference semantics), they reason that optional<T&> should be copy assignable as well.
In contrast, I believe it is more helpful to think of optional<T> as a container of T than as inheriting the interface of T; after all, optional<T> doesn't inherit other member functions from T, nor would it be correct to try and do so (despite it being suggested in N4173). Thus, I have disabled copy assignment for optional_ref<T>, as classes that contain T& are not copy assignable. The desire to implement reference semantic copy assignment may also be in part due to the misconception that the inability to rebind T& is a "flaw" in its design. On the contrary, the inability to rebind T& is a feature that is intrinsic to its nature (hence the apparent dilemma of value vs. reference semantics for copy assignment of a reference-like type). The behaviour of the optional<T&> suggested in the auxiliary proposal can be closely reproduced by optional<reference_wrapper<T>>.
Currently, nowhere in the guidelines is optional<T> mentioned. This is understandable, given that C++17 is still a work in progress. Presumably the guidelines will be updated when C++17 is released. Even then, it will be some time before everyone following the guidelines has access to a production-ready implementation of the C++17 standard library. Given the fundamental role that optional<T> plays in accurately representing the concept of an "optional" value—an extremely common requirement—I suggest adding an implementation of optional<T> to the GSL. Note that I still suggest adding optional_ref<T> as opposed to implementing optional<T&>, since any implementation of optional<T> should ideally be standard-conforming. If and when optional<T&> is standardized, the guidelines can be updated and instances of optional_ref<T> can easily be find-and-replaced (providing optional<T&> has roughly the same semantics as optional_ref<T>).
Introducing retained<T> and optional_ref<T> to the GSL and guidelines would not eliminate the need for pointer annotations such as owner<T> and not_null<T>. Annotations are necessary to help static analysis tools verify the integrity of low-level and legacy code that must use pointers for whatever reason. However, there are a number of things to be said about the current approach.
Pointer types, as explained in this document, are multi-purpose, and thus support far to broad a range of operations for any one use case:
- Array subscriptions makes sense only for pointers to arrays or elements of an array
- Pointer arithmetic operations make sense only for pointers to elements of an array
- Assigning and checking for null makes sense only for pointers that can be null
- Calling
deleteordelete[]makes sense only for pointers returned bynewornew[]
I suggest that a bare T* (i.e. one without any annotations) should only represent a non-owning pointer to a single object when pointers must be used. However, T* should also be implicitly "not null". This allows individual "features" of pointers to be enabled using annotations, in the vein of owner<T>, rather than the approach taken by not_null<T>, which is to disable a potentially unsafe "feature" that T* is assumed to have by default. Therefore, I suggest supporting the following annotations, one to enable each of the "features" listed above:
array<T>enables array subscriptioniterator<T>enables pointer arithmetic operations (and array subscription)nullable<T>enables the null pointer stateowner<T>enables use of eitherdeleteordelete[](when combined witharray<T>)
Note the distinction between owners of objects and owners of arrays, something the guidelines currently do not account for. I have designated array<T> for this purpose, rather than iterator<T>, as it would be error-prone to do something like p++ on an owner<iterator<T>>. An example of correct use of these annotations is:
T* p1 = &t;
array<T*> p2 = &arr;
nullable<owner<T*>> p3 = new (nothrow) T;
owner<T*> p4 = new T;
owner<array<T*>> p5 = new T[n];
iterator<T*> p6 = p5;
This approach would allow static analysis tools to enable a small "safe" subset of operations for T* by default, allowing specific "unsafe" features to be enabled one-by-one as required using annotations. Static analysis tools can warn wherever a "feature" is used without its corresponding annotation. For example:
T t1 = p1[1]; // warning: array subscription without `array` or `iterator`
p2++; // warning: pointer arithmetic without `iterator`
T t2 = *p3; // warning: dereferencing `nullable` without null check
if (p4) { … } // warning: checking for null without `nullable`
delete p5; // warning: calling `delete` with `array`
p6 = nullptr; // warning: setting to null without `nullable`
Of course, such warnings are likely to be ubiquitous in old C++ code, but there is really no way around this if your goal to make your code safer, more explicit and free of bugs. Static analysis tools could potentially implicitly annotate certain variables, especially in function scope. For example:
template <typename T, typename N>
T* find(T (&a)[N], T const& t) {
T* it = begin(a); // `it` is implicitly `iterator<T*>`
for (; it != end(a); ++it) { // `++it` is okay
if (*it == t) break;
}
return it; // warning: discarding `iterator`
}
This could automatically reduce the number of warnings by only flagging violations on API boundaries. In addition, static analysis tools could provide ways to disable particular categories of warning, or show warnings only for particular files or sections of code. Updating old code is messy business, but there are a variety of ways to make it more manageable.
The nullable<T> annotation, as a mere template type alias, lacks the ability to enforce the "not null" condition at run-time like not_null<T>. This may seem like a loss of functionality, but considering that debug builds are likely to be able to catch attempts to dereference null pointers at run-time, and that run-time checks will probably be turned off for release builds, along with the fact that use of T&, retained<T> and span<T> allows the "not null" condition to be enforced at compile-time, not_null<T> may actually provide little additional value.
One could argue that not_null<T> could be used in situations where switching to a higher-level abstraction would break too much client code. However, not_null<T> explicitly disables pointer arithmetic, which means that it already breaks code where T* is used as an iterator. In fact, it seems that not_null<T> is actually a cross between an annotation and a high-level type. It is simultaneously trying to facilitate the job of static analysis tools and itself perform safety checks at run-time and compile-time. In addition, the question arises, how would one annotate not_null<T> to verify that its implementation is correct?
Given that the introduction of retained<T> would all but replace use of not_null<T*> in high-level code, and that not_null<T*> seems to be inherently incompatible with smart pointer types, it seems reasonable to replace it with the nullable<T> annotation. It would not be necessary to remove not_null<T> from the GSL immediately, if there is concern that doing so would break a lot of existing code.
The aim of the C++ Core Guidelines is to facilitate writing of modern C++ code. At the core of modern C++ is type-safety, and pointers are not type-safe. Rather than inventing arbitrary rules about how T* should be used in modern, high-level C++ code, we should provide strongly typed alternatives to the old, unsafe uses of T*. In addition, a full range of type alias pointer annotations could be provided to allow the features of pointers to be enabled one-by-one wherever it is not possible to replace T* with higher-level abstractions. This would allow effective static validation of low-level and legacy code that still uses pointers, including standard library and GSL implementations.