Skip to content

CXX-3232 add bsoncxx v1 declarations #1412

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 14 commits into
base: master
Choose a base branch
from

Conversation

eramongodb
Copy link
Contributor

@eramongodb eramongodb commented May 28, 2025

Resolves CXX-3232. Followup to #1318, #1401, and #1402.

This is 2 out of an estimated 7 major PRs which in total are expected to resolve CXX-2745.

This PR introduces the very first set of v1 ABI exported symbols. 🎉

image

... only because MSVC complains otherwise. The rest will come as planned in CXX-3233 (impls).


Due to the volume of changes relative to the equivalent v_noabi interfaces, the bulk of relative changelog (v_noabi -> v1) and rationale for design decisions are documented inline as GitHub comments (coincidentally, exactly 100 comments in total).

In summary, notable changes include:

  • General Notes
    • Nearly all preconditions are now documented and well-defined to instead throw an exception or yield an "invalid" state.
    • Narrowing casts and discarded return values are now thoroughly accounted for as new error codes.
    • All potentially-thrown exceptions are now documented and include the list of error codes which may be thrown.
  • v1::error
    • All exceptions thrown by library components have a corresponding error code enumeration declared in v1::error with corresponding error category declared in v1::error::category.
    • Error conditions source and type are provided as the primary user-facing error code query mechanism when component-specific error codes are not necessary.
  • v1::types::b_*
    • All types are now consistently non-aggregate for list-initialization consistency.
    • All "single underlying value" types now consistently support implicit conversion to said value.
    • Multi-argument constructors are no longer explicit (always favored over value).
  • v1::types::view
    • This type is now semitrivial (no user-provided non-default special member functions).
    • Conversion to this type is now favored over conversion to value when applicable.
  • v1::types::value
    • All (non-default) constructors are now explicit (b_* is always favored) so either b_* or view (cheap) is always preferred to value (expensive) when applicable.
    • v1::types::id constructors are removed in favor of b_* and converting constructors.
    • "Inline PIMPL" avoids redundant allocation while preserving encapsulation.
  • v1::element::view
    • The "invalid" state is now clarified and thoroughly documented.
    • "Inline PIMPL" leaves room for an alternative bool-based approach to "missing field" diagnostics (CXX-2120).
    • Although not semitrivial, the class is still nothrow copyable.
    • Equality comparison is defined as "same field in same BSON binary data". Relational comparison is also possible but deferred for now.
  • v1::document::view
    • Now supports an "invalid" state without underlying BSON binary data.
  • v1::array::view
    • Now (nearly) completely defined in terms of v1::document::view.
    • API is extended for better parity with v1::document::view (CXX-2118).
  • v1::document::value
    • Now default-constructible (CXX-939), equivalent to an "invalid" v1::document::view.
    • deleter_type is now a std::function<T> instead of a function pointer.
    • Customization points for to_bson and from_bson are now properly constrained.
  • v1::array::value
    • Now (nearly) completely defined in terms of v1::document::value.
    • API is extended for better parity with v1::document::value (CXX-2118).

@eramongodb eramongodb self-assigned this May 28, 2025
@eramongodb eramongodb force-pushed the cxx-abi-v1-decls branch 2 times, most recently from b52a537 to 07f460e Compare May 30, 2025 17:52
if (("$ret" & 0x03)); then # ABIDIFF_ERROR (1) | ABIDIFF_USAGE_ERROR (2)
echo "abidiff error" >&2
exit 1
elif (("$ret" & 0x08)); then # ABIDIFF_ABI_INCOMPATIBLE_CHANGE (8).
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Per abidiff docs:

The exit code of the abidiff command is either 0 if the ABI of the binaries being compared are equal, or non-zero if they differ or if the tool encountered an error. In the later case, the exit code is a 8-bits-wide bit field in which each bit has a specific meaning.

This PR introduces the first exported v1 ABI symbol: bsoncxx::v1::exception::~exception(). This new symbol triggered a non-zero exit code of ABIDIFF_ABI_CHANGE (4), which was incorrectly handled as incompatibility/failure. The script is updated to distinguish the individual error codes appropriately: we only care about incompatible changes. Compatible changes (such as this) will still be included in the resulting abidiff report.

std::is_trivially_copy_constructible<T>,
std::is_trivially_copy_assignable<T>> {};
struct is_semitrivial
// https://gcc.gnu.org/onlinedocs/gcc-4.9.0/libstdc++/manual/manual/status.html
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Unfortunately, GCC 4.X (used by RHEL 7) does not support the std::is_trivially_* type traits. Special-case this compiler version by simply assuming this property to be true (we have plenty of true coverage with other compilers and compiler versions on other platforms).

Comment on lines +29 to +31
static_assert(
is_nothrow_moveable<view::const_iterator>::value,
"bsoncxx::v1::document::view::const_iterator must be nothrow moveable");
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Despite v1::document::view being semitrivial, its iterator is only nothrow moveable due to its internal v1::element::view data member.

Comment on lines +29 to +31
static_assert(
is_nothrow_moveable<view::const_iterator>::value,
"bsoncxx::v1::array::view::const_iterator must be nothrow moveable");
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Despite v1::array::view being semitrivial, its iterator is only nothrow moveable due to its internal v1::element::view data member.

namespace bsoncxx {
namespace v1 {

exception::~exception() = default;
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is the key function for the bsoncxx::v1::exception class and the first v1 ABI symbol (along with the vtable) to be exported by the bsoncxx library. 🎉

Although this is meant a declarations-only PR, this definition is required to support successful builds with MSVC, as otherwise:

error LNK2019: unresolved external symbol "public: virtual __cdecl bsoncxx::v1::exception::~exception(void)" (??1exception@v1@bsoncxx@@UEAA@XZ) referenced in function "public: virtual void * __cdecl bsoncxx::v1::exception::`vector deleting destructor'(unsigned int)" (??_Eexception@v1@bsoncxx@@UEAAPEAXI@Z)

Comment on lines +276 to +278
deleter_type const& get_deleter() const {
return _data.get_deleter();
}
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

A non-const .get_deleter() is deliberately not provided to prevent "desync" of the deleter from the associated BSON binary data. Read-only access is nevertheless essential to support inspection of the type of underlying deleter, such as for v_noabi compatibility (e.g. "Is this deleter compatible with v_noabi::document::value::deleter_type?"). This will be explored in CXX-3234 (v_noabi refactor).

Comment on lines +314 to +319
///
/// Implicitly convert to `this->view()`.
///
/* explicit(false) */ operator v1::document::view() const {
return this->view();
}
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Asymmetry: value -> view is implicit, but view -> value is explicit.

Comment on lines +317 to +322
///
/// Implicitly convert to `this->view()`.
///
/* explicit(false) */ operator v1::types::view() const {
return this->view();
}
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Asymmetry: value -> view is implicit, but view -> value is explicit.

Comment on lines 307 to 312
///
/// Return a view of the BSON binary data as a document.
///
v1::document::view view() const {
return {_data.get(), _length};
}
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

All member functions below are defined in terms of v1::document::view, including error codes and Doxygen documentation.

Important

The new "invalid" state of a v1::document::view allows this conversion to be well-defined even when a value is default-initialized (_data.get() == nullptr).

class value {};
class value {
private:
v1::document::value _value;
Copy link
Contributor Author

@eramongodb eramongodb May 30, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

All behavior for this class is defined in terms of v1::document::value or v1::array::view, including error codes and Doxygen documentation.

@eramongodb eramongodb marked this pull request as ready for review May 30, 2025 20:31
@eramongodb eramongodb requested a review from a team as a code owner May 30, 2025 20:31
@eramongodb eramongodb requested review from rcsanchez97, kevinAlbs and vector-of-bool and removed request for rcsanchez97 May 30, 2025 20:32
@eramongodb eramongodb changed the title CXX-3232 add v1 interface declarations CXX-3232 add bsoncxx v1 declarations May 30, 2025
Comment on lines +248 to +250
/// @note This iterator almost satisfies Cpp17ForwardIterator, but `std::iterator_traits<T>::reference` is defined as
/// `value_type`, similar to `std::vector<bool>::iterator` and `std::istreambuf_iterator<T>`. Therefore, this iterator
/// only fully satisfies Cpp17InputIterator.
Copy link
Contributor Author

@eramongodb eramongodb Jun 2, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Update: renamed Legacy -> Cpp17 for consistency with standard spec for C++20 and newer (by P0896R4; cppreference still uses "Legacy", a leftover from when the C++20 spec was still being drafted) + restoring this comment due to the associated changes of the prior comment being lost by GitHub.


The v1 ABI interfaces address and clarify several issues with iterators that were identified during work on CXX-3082:

  • Missing const qualifiers on multiple member functions, e.g. auto const iter = doc.begin(); use(*iter); is ill-formed (?!).
  • iterator_category = std::forward_iterator_tag despite not actually satisfying the requirements of LegacyForwardIterator:
    • e.g. does not satisfy it: const_iterator -> *it -> T const&.
    • e.g. does not satisfy a == b -> std::addressof(*a) == std::addressof(*b).
  • Operations on an end iterator behavior, such as deference yielding an invalid element and increment yielding another end iterator, are underdocumented not documented at all.

v1::document::view::const_iterator (and v1::array::view::const_iterator) addresses all these issues by ensuring it properly satisfies and declares itself to be a LegacyInputIterator instead (even though it may provide some features of a LegacyForwardIterator).

Copy link
Collaborator

@kevinAlbs kevinAlbs left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The bsoncxx ABI v1 interfaces look well researched, and I really like the efforts towards better defining expected exceptions / minimizing preconditions.

}

/// @copydoc v1::document::value::reset(v1::document::view const& v)
void reset(v1::array::view v) {
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

document::value::reset accepts a view const& but array::value::reset accepts a view. Consider making consistent.

Suggest using a non-reference view to avoid a "view to a view" type unless there is a benefit to accepting a view const&.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

unless there is a benefit to accepting a view const&.

I had meant to use view const& as a simplification of CppCoreGuidelines' "F.16: For “in” parameters, pass cheaply-copied types by value and others by reference to const" to avoid complications with determining what constitutes "cheaply-copied" (see also: here, here). However, it seems I was unable to apply this consistently. I am in favor of reverting to applying the cheap-to-copy guidelines as suggested.

For the v1 API, I believe the following requirements should suffice to define a "cheap-to-copy" type T:

  • sizeof(T) <= 2u * sizeof(void*) (common platform-independent heuristic for "fits in a register").
  • is_trivially_copyable<T> + is_trivially_destructible<T> (per the is_semitrivial<T> static assertions).

Given these conditions, v1::document::view and v1::array::view may be considered "cheap-to-copy" and thus passed-by-value for in parameters. The other view types are disqualified due to:

  • v1::element::view: too big + not trivially copyable or destructible.
  • v1::types::view: too big.

/// @par Includes
/// - @ref bsoncxx/v1/array/view.hpp
/// - @ref bsoncxx/v1/document/value.hpp
/// - @ref bsoncxx/v1/element/view.hpp
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Add include of element/view.hpp?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

v1/element/view.hpp is (documented as) provided by both v1/array/view and v1/document/value, so I will opt to remove this inclusion for consistency with the current set of #include directives.

/// - @ref data() is not null, and
/// - @ref size() is not less than `5` (the minimum size of a BSON document).
///
/// The BSON binary data being represented is only validated as minimally required to satisfy a requested operation.
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
/// The BSON binary data being represented is only validated as minimally required to satisfy a requested operation.
/// The BSON data being represented is only validated as minimally required to satisfy a requested operation.

Suggest removing "binary" to avoid mixing with the BSON binary type.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Reworded "BSON binary data" -> "BSON bytes" to continue ensuring that the term is being used to refer to the (raw) byte representation of the BSON type.

///
/// Return a const iterator to the element within the represented BSON array at index `i`.
///
/// If this view is invalid, returns an end iterator.
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggest documenting expectation if i is not found.

Also suggest documenting this is not a constant time operation. That may help warn users wanting to iterate over indexes like:

for (uint32_t i = 0; i < num_entries; i++) {
    sum += arr.view()[i].get_int32(); // Scans BSON each time!
}

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Added a "Complexity" paragraph mirroring the standard library's "Complexity" specification (alongside "Preconditions", "Constriants", etc.).

///
/// Zero-initialize the byte representation.
///
/// @note The result is equivalent to `"0E-6176"`, not `"0"`.
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I would rather avoid a subtle quiet behavior change between v_noabi and v1. LGTM as proposed.

///
/// Initialize with `code` and `scope`.
///
b_codewscope(v1::stdx::string_view code, v1::document::view scope) : code{code}, scope{scope} {}
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

consider adding a code-only constructor

I support leaving as-is as a minor simplification since the code-with-scope BSON type is deprecated.

///
/// Initialize with `increment` and `timestamp`.
///
b_timestamp(std::uint32_t increment, std::uint32_t timestamp) : increment{increment}, timestamp{timestamp} {}
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

consider adding a increment-only constructor

I support leaving it out (as proposed) unless there is a known need.


public:
///
/// Initialize as an empty view.
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Confirming: is v1::document::view initialized to an empty document intended for consistency with v_noabi::document::view?

If there were no precedent, I might rather the initialize to an invalid, but I expect that is a quiet behavior change that might be missed when migrating. So I support this even if it means a slight asymmetry between default constructed v1::document::value (invalid) and default constructed v1::document::view (invalid).

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It is both for compatibility with v_noabi::document::view as well as consistency with std::string_view, which is default-initialized as an "empty string" rather than an "invalid string view".

It follows that for consistency with std::string, a default-constructed v1::document::value should also be an empty document rather than invalid. However, as noted in CXX-939 and related discussions, this complicates the allocation strategy for representing an empty BSON document (uncondition allocation? use pre-allocated storage? etc.).

We could consider leaving the currently-proposed v1::document::value semantics the same, but make the following tweak to the behavior of this->view():

v1::document::view view() const {
    if (v1::document::view const v{_data.get(), _length}) {
        return v;
    }
    return {};
}

The result of this tweak would be that a default-initialized value is equivalent to a default-initialized view (both an "empty document") even when value._data.get() == nullptr. No unconditional or special-cased allocation is required. Since all non-ownership-related accessors are implemented in terms of this->view(), the null pointer would not be "observable" to the user via normal BSON document operations except via this->release(). However, this PR did not initially propose this due concern that this may be more confusing to users than the current inconsistency between view vs. value default initialization.

Thoughts?

Comment on lines 1150 to 1151
/// When the type of `lhs` and `rhs` compare equal but are not a supported value (not defined by @ref
/// BSONCXX_V1_TYPES_XMACRO), the result is always `true`.
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Instead, all v1 interfaces ensure equality comparison has no preconditions and cannot fail.

Confirming: is avoiding an exception intended to avoid preconditions in general? Or is equality treated specially to avoid throwing?

///
/// The number of bytes required to represent an ObjectID.
///
static constexpr BSONCXX_ABI_EXPORT std::size_t k_oid_length = 12;
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No strong opinion. I am OK with the proposed change (appears consistent with v_noabi), but would not object to switching to an enum constant to avoid adding to the ABI.

@vector-of-bool do you have an opinion?

Copy link
Contributor Author

@eramongodb eramongodb left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggestions applied. Open questions following first round of review include:

  • Should compile-time constants be defined as enum constants (not exported in the ABI) instead of static constexpr variables (exported in the ABI due to pre-C++17 compatibility requirements)?
  • Should v1::document::value::view() return an "empty doc" view instead of an invalid view without otherwise changing its allocation behavior (_data.get() == nullptr)?
  • Should v1::document::value continue to support implicit assignment from a v1::document::view (currently requires doc = value{view};)?

}

/// @copydoc v1::document::value::reset(v1::document::view const& v)
void reset(v1::array::view v) {
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

unless there is a benefit to accepting a view const&.

I had meant to use view const& as a simplification of CppCoreGuidelines' "F.16: For “in” parameters, pass cheaply-copied types by value and others by reference to const" to avoid complications with determining what constitutes "cheaply-copied" (see also: here, here). However, it seems I was unable to apply this consistently. I am in favor of reverting to applying the cheap-to-copy guidelines as suggested.

For the v1 API, I believe the following requirements should suffice to define a "cheap-to-copy" type T:

  • sizeof(T) <= 2u * sizeof(void*) (common platform-independent heuristic for "fits in a register").
  • is_trivially_copyable<T> + is_trivially_destructible<T> (per the is_semitrivial<T> static assertions).

Given these conditions, v1::document::view and v1::array::view may be considered "cheap-to-copy" and thus passed-by-value for in parameters. The other view types are disqualified due to:

  • v1::element::view: too big + not trivially copyable or destructible.
  • v1::types::view: too big.

/// @par Includes
/// - @ref bsoncxx/v1/array/view.hpp
/// - @ref bsoncxx/v1/document/value.hpp
/// - @ref bsoncxx/v1/element/view.hpp
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

v1/element/view.hpp is (documented as) provided by both v1/array/view and v1/document/value, so I will opt to remove this inclusion for consistency with the current set of #include directives.

/// - @ref data() is not null, and
/// - @ref size() is not less than `5` (the minimum size of a BSON document).
///
/// The BSON binary data being represented is only validated as minimally required to satisfy a requested operation.
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Reworded "BSON binary data" -> "BSON bytes" to continue ensuring that the term is being used to refer to the (raw) byte representation of the BSON type.

///
/// Return a const iterator to the element within the represented BSON array at index `i`.
///
/// If this view is invalid, returns an end iterator.
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Added a "Complexity" paragraph mirroring the standard library's "Complexity" specification (alongside "Preconditions", "Constriants", etc.).

Comment on lines 115 to 116
/// b = a["2"]; // If "a.b" exists and is a document, `b` is valid. Otherwise, `b` is invalid.
/// c = b[3]; // If "a.b.c" exists and is an array, `c` is valid. Otherwise, `c` is invalid.
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thank you for the suggestions. These are indeed more accurate.

Comment on lines 286 to 292
friend bool operator==(view const& lhs, view const& rhs) {
if (!lhs != !rhs) {
return false;
}

return !lhs || (lhs.raw() == rhs.raw() && lhs.offset() == rhs.offset());
}
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Makes sense. Reverted v1::element::view to be non-equality-comparable.

The operation is moved back into const_iterator's equality comparison operator.


public:
///
/// Initialize as an empty view.
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It is both for compatibility with v_noabi::document::view as well as consistency with std::string_view, which is default-initialized as an "empty string" rather than an "invalid string view".

It follows that for consistency with std::string, a default-constructed v1::document::value should also be an empty document rather than invalid. However, as noted in CXX-939 and related discussions, this complicates the allocation strategy for representing an empty BSON document (uncondition allocation? use pre-allocated storage? etc.).

We could consider leaving the currently-proposed v1::document::value semantics the same, but make the following tweak to the behavior of this->view():

v1::document::view view() const {
    if (v1::document::view const v{_data.get(), _length}) {
        return v;
    }
    return {};
}

The result of this tweak would be that a default-initialized value is equivalent to a default-initialized view (both an "empty document") even when value._data.get() == nullptr. No unconditional or special-cased allocation is required. Since all non-ownership-related accessors are implemented in terms of this->view(), the null pointer would not be "observable" to the user via normal BSON document operations except via this->release(). However, this PR did not initially propose this due concern that this may be more confusing to users than the current inconsistency between view vs. value default initialization.

Thoughts?

Comment on lines 1150 to 1151
/// When the type of `lhs` and `rhs` compare equal but are not a supported value (not defined by @ref
/// BSONCXX_V1_TYPES_XMACRO), the result is always `true`.
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The latter. Even if they may not be noexcept, it is usually astonishing to the user when the expression a == b can "fail". It is difficult to justify an equality comparison for a regular type having preconditions (that is, possibly UB) or to throw an exception (the postconditions could not be satisfied): that is, why should either be the case? There is an argument for comparison with "unsupported values" to throw an exception, but I believe "unspecified" is a safer and more user-friendly approach.

On that note, took this opportunity to change the documentation of comparing unsupported types from "true" to "unspecified".

Comment on lines +219 to +242
///
/// Equivalent to `to_bson(v, *this);` after default-initialization.
///
/// @par Constraints:
/// - `T` is not @ref bsoncxx::v1::document::value.
/// - `to_bson(v, *this)` is a valid and unambiguous overload found by ADL.
///
template <typename T, detail::enable_if_t<has_to_bson<T>::value>* = nullptr>
explicit value(T const& v) : value{} {
to_bson(v, *this); // ADL.
}

///
/// Equivalent to `*this = value{v}`.
///
/// @par Constraints:
/// - `T` is not @ref bsoncxx::v1::document::value.
/// - `to_bson(v, *this)` is a valid and unambiguous overload found by ADL.
///
template <typename T, detail::enable_if_t<has_to_bson<T>::value>* = nullptr>
value& operator=(T const& v) {
*this = value{v};
return *this;
}
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

After further investigation, it seems that default constructibility is possible to support while preserving backward compatibility with v_noabi::document::value by making the default constructor explicit:

explicit value() = default;

This ensures that existing patterns of the form v_noabi::document::value doc({}); to workaround lack of default-constructibility continue to work as expected without leading to ambiguity with SMFs: only the view constructor's parameter can be copy-initialized with {}.

The aforementioned issues concerning the lack of constraints was confusing the default constructibility problem with view -> value implicit conversion on assignment, where v_noabi::document::value doc; doc = view; is permitted not by an operator=(view), but via operator=(T const&) [T = view] despite no to_bson overload.

An operator=(view) could be added for consistency with v_noabi, but the current API proposes omitting this overload to ensure consistency of view -> value conversion being explicit.

Thoughts?

@kevinAlbs kevinAlbs self-requested a review June 26, 2025 19:21
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants