P2013R4
Freestanding Language: Optional ::operator new

1. Revision History

1.1. R4

Rebasing to N4878. Accounting for shifted numbering in [new.delete].

1.2. R3

In the design, an ODR-use without a definition no longer requires ill-formedness. The intent is still the same, in that it typically results in a linker error.

Updated grammar and ill-formedness of wording.

Found an additional note in expr.new/10 that needed adjustment.

Deferred feature test macro decision to a future revision of P2198.

1.3. R2

Added words in the design to indicate that the declarations in <new> are not changing.

Mentioning that pointer safety remains unchanged.

1.4. R1

Added discussion on feature test macro.

Added discussion on not fixing virtual destructors.

Declaring that constexpr new must continue to work.

Added wording.

Wording includes changes to library clauses, so added LEWG to the audience.

1.5. R0

The proposed solution is different than the one proposed for P1105R1, but the motivation is the same. The solution from P1105R1 is still listed as a design alternative.

2. What is changing

As a consequence of the above, runtime coroutines on freestanding implementations that are relying on the global allocation functions will also have implementation defined behavior, so long as the vendor supplied implementation is being used.

No other core language features require ::operator new. basic.stc.dynamic.allocation

::operator delete will be implementable as a no-op function on implementations that have do-nothing or undefined-behavior ::operator new implementations.

3. What is staying the same

Calling ::operator delete on a non-null pointer that did not come from ::operator new is still undefined behavior new.delete.single new.delete.array. Calling delete on an object or base that didn’t come from new is still undefined behavior expr.delete. This is what makes a no-op ::operator delete a valid strategy for freestanding vendor supplied ::operator new implementations.

The replaceable allocation functions will still be implicitly declared at global scope in each translation unit basic.stc.dynamic.general. A definition of the replaceable allocation functions must still exist. Non-ODR-uses of the replaceable allocation functions are still permitted (e.g. inside of uninstantiated templates). The declarations of ::operator new in the <new> header are still required to be present. Implementations of the replaceable allocation functions can be performed by linking in an extra translation-unit with the definitions of the functions.

"constexpr new" is still required to work, even when the vendor supplied replaceable allocation functions have not been replaced. The calls to ::operator new are required to be omitted, so any undefined behavior in the ::operator new implementation will also be skipped.

Non-allocating placement ::operator new (colloquially "placement new") and ::operator delete (colloquially "placement delete") are still required to be present in freestanding implementations.

Core language concepts of pointer safety remain unchanged. Note that the pointer safety library facilities util.dynamic.safety are not required to be present in freestanding implementations, and the author is not aware of any papers to make the pointer safety library facilities required in freestanding implementations.

Hosted implementations are unchanged. Users of freestanding implementations can still provide implementations of the replaceable allocation and deallocation functions. The behavior of virtual destructors is unchanged. The behavior of class specific operator new and operator delete overloads is unchanged. The requirements on user-provided ::operator new and ::operator delete overloads remains the same, particularly those requirements involving error behaviors. Coroutines will behave the same so long as promise-specific allocators are used. The storage for exception objects will remain unspecified.

4. Why?

4.1. No allocations allowed

FreeRTOS allows for both static and dynamic allocation of OS constructs [FreeRTOS_StaticVDynamic]. Static allocation in conjunction with ::operator new diagnostics can help avoid overhead and eliminate accidental usage.

THREADX [THREADX] does not consider dynamic allocation a core service, and can be built without support for dynamic allocation in order to reduce application size. THREADX also distinguishes between byte allocation (general purpose) vs. block allocation (no-fragmentation elements of fixed size in a pool).

Also, by allowing a no-op ::operator delete implementation, these space constrained applications can save code-size. No code needs to be present for ::operator delete synchronization, free block coalescing, or free block searching.

4.2. No right way to allocate memory

As an example, in the Microsoft Windows kernel environment, there are two leading choices about where to get dynamic memory [MSPools]. Users can get memory from the non-paged pool, which is a safe, but scarce resource; or users can get memory from the paged pool, which is plentiful, but not accessible in many common kernel operations. Non-paged pool must be used any time the allocated memory needs to be accessible from an interrupt or from a "high IRQL" context. The author has had experience with both paged pool and non-paged pool as defaults, with the predictable outcome of crashes with paged pool defaults and OOM with non-paged pool defaults. The implementer of the C++ toolchain is not in a good position to make this choice for the user.

In the Linux kernel environment, kmalloc [kmalloc] with the GFP_KERNEL flag should be used when allocating memory within the context of a process and outside of a lock, but the GFP_ATOMIC flag should be used when allocating memory outside the context of a process, such as inside of an interrupt. The implementers of the C++ runtime are in no position to know which is the correct flag to use by default. Using GFP_KERNEL when GFP_ATOMIC is needed will result in crashes from interrupt code and deadlocks. Using GFP_ATOMIC when GFP_KERNEL is appropriate will result in reduced system performance, spurious OOM errors, and premature exhaustion of emergency memory pools.

Freestanding implementations are intended to run without the benefit of an operating system (intro.compliance). However, the name of the function that supplies dynamic memory is usually an OS-specific detail. The C++ implementation should not (and may not) know the name of the function to request memory. The Windows kernel uses ExAllocatePoolWithTag. In the Linux kernel, kmalloc is the main function to use. In FreeBSD, a function named malloc is present, but it takes different arguments than the C standard library function of the same name. FreeRTOS uses pvPortMalloc, and THREADX uses tx_byte_allocate. Home-grown OSes will likely have other spellings for memory allocation routines.

Today’s C++ implementations don’t provide ::operator new implementations for all possible targets. Doing so isn’t a plausible goal, especially when the home-grown OSes are taken into account. This means that users are already forced into choosing between not having ::operator new support and providing their own implementation. We should acknowledge and standardize this existing practice, especially since we already have the extension point mechanism in place.

4.3. What about allocators?

Allocators are still useful in a less-than-minimal freestanding implementation. In environments with dynamic memory, custom allocators can be written and used with standard containers, assuming that the containers are present in the implementation. This could be done even if a global ::operator new is not present. The author has used stlport::vector<int, PageLockedAllocator> successfully in these environments.

std::allocator is implemented in terms of global ::operator new. In practice, it would be easy for an implementation to have an implementation of std::allocator in a header / module, and have that header still compile just fine. If the user has provided a global ::operator new, then std::allocator would have the same semantics as mandated for hosted implementations. If the global ::operator new is vendor supplied, then uses of std::allocator would invoke implementation defined behavior, and hopefully cause a diagnostic.

Some facilities in the standard library (e.g. make_unique) are implemented in terms of new, and not an allocator interface. It is useful to make these facilities generate diagnostics when dynamic memory isn’t available, and it is also useful to be able to control which memory pool is used by default.

4.4. virtual destructors

Ideally, if neither new nor delete is ever called, we wouldn’t need an operator delete. This proposal still requires some operator delete to exist, though that operator delete can be a no-op.

4.5. Why not fix virtual destructors, instead of keeping a no-op operator delete?

4.6. Feature test macro

In order to provide this macro, library implementations are going to require knowledge of the target environment. That knowledge may be via a list of known target platforms that are detected at build time, or by having the builder of the implementation supply that information in a configuration flag.

The most likely usage of a feature test macro for this feature is to conditionally define a custom ::operator new iff the implementation did not provide one by default. This is dangerous territory, as it encourages libraries to provide the one-and-only ::operator new definition. If two such libraries do this, then there is an ODR issue.

Another likely usage is to fall back to an implementation that does not use the heap at all.

#if defined(__cpp_lib_no_default_operator_new) && __cpp_lib_no_default_operator_new >= 20200913
  using my_container = fixed_capacity_vector;
#else
  using my_container = my_vector;
#endif

This is an imprecise check. Even though there is no default operator new, there may be a user provided operator new that works fine.

If this feature test macro were provided in the positive (__cpp_lib_has_default_operator_new), it wouldn’t be useful for a very long time.

#if defined(__cpp_lib_has_default_operator_new) && __cpp_lib_has_default_operator_new >= 20200913
  using my_container = my_vector;
#else
  // spuriously triggers for C++20 and earlier code
  using my_container = fixed_capacity_vector;
#endif

If this feature test macro were provided in the negative (__cpp_lib_no_default_operator_new), it would be the only feature test macro that wouldn’t be required to be defined in a conforming implementation.

All these considerations are deserving of dedicated SG10 discussion with P2198, and should not hold up the progress of this paper.

4.7. Likely misuses and abuses

5. Experience

6. Polling history

6.1. Aug 19, 2020, EWG, Telecon

SF/F/N/A/SA

2/16/1/0/0

6.2. Feb 14, 2020, EWG, Prague

SF/F/N/A/SA

8/10/7/0/0

6.3. Feb 12, 2020, EWGI, Prague

SF/F/N/A/SA

7/6/1/0/0

Is a Feature test macro a valuable addition to this paper?

SF/F/N/A/SA

0/5/5/2/0

Do we believe that P2013 is sufficiently developed to be seen by EWG?

SF/F/N/A/SA

5/8/0/0/0

6.4. Jan 8, 2020 SG14 Telecon

SF/F/N/A/SA

9/10/0/0/0

approves to go to EWG

7. Design Alternatives

7.1. Alternative 0: Implementation defined allocating ::operator new (Proposed above)

7.2. Alternative 1: Optional throwing ::operator news, no-op default deallocation functions

The nothrow_t overloads are specified to forward to an appropriate throwing overload. That implementation would still be fine on a system without dynamic storage available. This alternative was not selected as it is more difficult to teach, and because the target audience would likely be astonished that the nothrow_t overload has a try/catch in it.

7.3. Alternative 2: No deallocation functions

This alternative has the benefit of being zero overhead and very explicit, but it has troublesome consequences for implementations. There are several language support classes that have virtual destructors, and something would need to be decided for them. Notably, type_info and the exception hierarchy all have virtual destructors. The standard library implementers may be prohibited from providing operator new and operator delete overloads (conforming#member.functions). Alternatively, the facilities that require classes with virtual destructors could all be off-limits until operator delete was made available. This would eliminate many cases with exceptions, dynamic_cast on references, and typeid.

If we were to adopt this alternative, many users would provide a no-op ::operator delete in their code, giving their code the same semantics and trade-offs as the proposed solution.

7.3.1. Experience

7.4. Alternative 3: No deallocation functions and new ODR-used rules for virtual destructors

7.4.1. How could this virtual destructor ODR-use change be implemented?

Existing linkers already have the ability to take multiple identical virtual table implementations and pick one for use in the final binary. A potential implementation strategy is for compilers and linkers to support a new "weaker" linkage. When the default heap is disabled, the compiler would emit a vtable with a nullptr or pure virtual function in the virtual destructor slot. When new is called, a "stronger" linkage vtable would be emitted that has the deleting destructor in the virtual destructor slot. The linker would then select a vtable with the strongest linkage available. Today’s linkage would be considered "stronger". Only partially filled vtables would have "weaker" linkage.

7.4.2. ABI impact

Shared libraries are trickier. Vtables aren’t always emitted into every translation unit. Take shared library "leaf" that has a default heap. It depends upon shared library "root" that does not have a default heap. If a class with a virtual destructor is defined in "root", along with its "key function", then a call to new on the class in "leaf" will generate an object with a partial vtable. Calling delete on that object will cause UB (usually crashes).

Lack of a default heap should generally be considered a trait of the platform. Mixing this configuration shouldn’t be a common occurrence.

7.4.3. Experience

7.5. Alternative 4: Missing ::operator new, no-op default deallocation functions (Proposed in R3)

This approach is substantially more difficult with constexpr new though. Colloquially, users would expect that uses of ::operator new at compile time would not require a runtime definition of ::operator new. However, constexpr doesn’t work like that. Implementations are permitted to emit a runtime definition of constexpr functions, as well as keep an internal representation for compile time evaluation, even if the function only happens to be used for constant evaluation. The runtime definition could still cause linker errors, even if nothing is calling the function.

It is possible to rework the ODR rules such that a runtime definition is only emitted if needed. That is a substantially larger change though, and there was resistance from CWG to make this large of a change to ODR at this time. This approach should be reconsidered for adoption if more OS-dependent facilities become available at compile in the future (e.g. constexpr iostreams). A lower effort approach is acceptable for now though, since ::operator new and ::operator delete are the only such facilities at present.

8. Wording

8.1. new.delete

17.6.3 Storage allocation and deallocation [new.delete]

17.6.3.1 General [new.delete.general]

1 Except where otherwise specified, the provisions of 6.7.5.5 apply to the library versions of operator new and operator delete. If the value of an alignment argument passed to any of these functions is not a valid alignment value, the behavior is undefined.

2 On freestanding implementations, it is implementation defined whether the behavior of the implementation-provided replaceable global allocation functions satisfy the required behaviors described in [new.delete.single] and [new.delete.array].

[Note: A freestanding implementation can provide replaceable global allocation functions that cause undefined behavior. During constant evaluation, the behavior of the implementation-provided replaceable global allocation functions is irrelevant, as those calls are omitted ([expr.new]).- end note]

Recommended practice: If any of the implementation-provided replaceable global allocation functions meet the requirements of a hosted implementation, they all should.

9. Acknowledgments

Thank you to Daveed Vandevoorde and Walter Brown for providing feedback on the wording.