--- title: New reflection metafunction - is_structural_type (US NB comment 49) document: P3856R8 date: 2026-03-27 audience: LEWG/LWG reply-to: - "Jagrut Dave " - "Alisdair Meredith " paper-type: proposal --- ## Document #: P3856R8 [[Latest](https://wg21.link/P3856)] [[Status](https://github.com/cplusplus/papers/issues/2460)] Date: 2026-03-27 Project: Programming Language C++ Audience: LEWG/LWG Reply-to: Jagrut Dave ([jdave12@bloomberg.net](mailto:jdave12@bloomberg.net)) Alisdair Meredith ([ameredith1@bloomberg.net](mailto:ameredith1@bloomberg.net)) Contents 1. Revision History 2. Introduction 3. Structural types 3.1 is_structural_type ### 3.1.1 Motivating examples ### 1. Revision History From version 7 - Improved wording based on LWG - Fixed headings From version 5 - Added feature test macros From version 4 - Improved wording based on Barry’s feedback From version 3 - Removed *is_destructurable_type* and its sample implementation, per LEWG feedback - Removed the access context parameter from *is_structural_type* signature - Proposed type trait for *is_structural_type,* per LEWG feedback From version 2 - Corrected version number From version 1 - Added metafunction is_structural_type addressing US NB comment 49 From version 0 - Corrected paper number, added observations from the proof of concept (section 3.3) ### 2. Introduction Several parts of the standard and library refer to structural types, including library mandates that types be structural, yet there is no way to query whether a type is structural. Library mandates clauses mean that library implementers must somehow have this functionality, but it is simply not exposed to users. We propose a new reflection metafunction - *is_structural_type(info)* as a solution, which would return *true* when *info* is the reflection of a structural type, and *false* otherwise. We also propose a corresponding type trait. The reflection metafunction is proposed to be added to the ** header, and the type trait to . We provide a sample implementation using other reflection metafunctions from P2996 to demonstrate how far we can go with the existing batch of reflection type queries included in C++26. In addition, we compare the approaches of using traditional type traits vs. using reflection metafunctions to query type attributes. ### 3.1 is_structural_type As per [[temp.param] (13.2)](https://eel.is/c++draft/temp.param#12), a structural type is (a) a scalar type, (b) an lvalue reference type, or (c) a literal class type whose direct bases and non‑static data members are all public and non‑mutable, and whose types are themselves structural (or arrays thereof). Captureless lambdas are also considered structural, per [[expr.prim.lambda.closure] (7.5.6.2)](https://eel.is/c++draft/expr.prim.lambda.closure#4). The *is_structural_type check* is valuable wherever values (including small user-defined structs) have to be treated as compile‑time constants, since NTTPs must be structural types. E.g., better diagnostics can be produced for debugging template instantiations, and APIs can be cleanly separated. ```cpp template requires is_structural_v struct const_wrapper { static constexpr auto value = V; }; template requires (!is_structural_v) struct const_wrapper; // = delete; // or static_assert with a helpful note template requires is_structural_v void register_token() { /* Register type V */ } template void register_token(type_tag) { /* Register type_tag value */ } ``` A single logging API could be used for compile and run time logging. ```cpp template requires is_structural_type_v inline void log_event() { /* Tag is known to be NTTP */ } // Fall back, run-time. inline void log_event(std::string_view tag); ``` The code below shows a possible implementation of *is_structural_type,* using Bloomberg’s Clang fork. ```cpp #include namespace proposed { consteval bool is_structural_type(std::meta::info); } namespace detail { // Strip top-level cv, but not references (rvalue refs are *not* structural). consteval std::meta::info canonicalize(std::meta::info t) { return remove_cv(dealias(t)); } // Peel one array extent; P2996 has array transforms in consteval std::meta::info remove_one_extent(std::meta::info t) { return remove_extent(t); } consteval bool is_array_of_structural_types(std::meta::info t) { t = canonicalize(t); if (!is_array_type(t)) return false; // Recursively strip all extents, then check the element type while (is_array_type(elem)) { elem = remove_one_extent(elem); } return proposed::is_structural_type(canonicalize(elem)); } // Return false if any direct base is not public *or* its type isn’t structural. consteval bool bases_ok(std::meta::info t) { constexpr auto unchecked = std::meta::access_context::unchecked(); for (std::meta::info rel : bases_of(t, unchecked)) { if (!is_public(rel)) return false; // declared access std::meta::info rel_type = type_of(rel); // recursive check if (!proposed::is_structural_type(canonicalize(rel_type))) return false; return true; // Return false if any nsdm is not public or is mutable, or its type isn’t consteval bool data_members_ok(std::meta::info t) { constexpr auto unchecked = std::meta::access_context::unchecked(); for (std::meta::info m : nonstatic_data_members_of(t, unchecked)) { if(!is_accessible(m, unchecked)) return false; if (!is_public(m)) return false; if (is_mutable_member(m)) return false; std::meta::info MT = type_of(m); if (is_array_type(MT)) { if (!is_array_of_structural_types(MT)) return false; } else { if (!proposed::is_structural_type(canonicalize(MT))) return false; } } return true; } // Forward declaration for mutual recursion. consteval bool is_literal_type(std::meta::info type, std::meta::access_context ctx); // Helper: at least one variant member of a union is a non-volatile literal type. // (Empty unions are allowed by DR 2598.) consteval bool union_has_literal_variant(std::meta::info union_type, auto ms = nonstatic_data_members_of(union_type, ctx); if (ms.empty()) return true; // union with 0 variant members for (auto m : ms) { if(!is_accessible(m, ctx)) return false; std::meta::info mt = dealias(type_of(m)); if (!is_volatile_type(mt) && is_literal_type(mt, ctx)) { return true; } return false; } // Helper: check “all bases and non-static data members are non-volatile literal types”, // and enforce the “anonymous union has a literal variant” rule for classes with anon unions. consteval bool class_members_are_literal(std::meta::info class_type, // Bases for (auto b : bases_of(class_type, ctx)) { std::meta::info bt = dealias(type_of(b)); if (is_volatile_type(bt) || !is_literal_type(bt, ctx)) { return false; } } // Non-static data members for (auto m : nonstatic_data_members_of(class_type, ctx)) { if(!is_accessible(m, ctx)) return false; std::meta::info mt = dealias(type_of(m)); // Anonymous union members need a special check (P2996 exposes has_identifier + union test). if (is_union_type(mt) && !has_identifier(m)) { if (!union_has_literal_variant(mt, ctx)) { return false; } continue; if (is_volatile_type(mt) || !is_literal_type(mt, ctx)) { return false; } return true; substitution. template constexpr bool __has_constexpr_dtor_if_default = false; template constexpr bool __has_constexpr_dtor_if_default< T, // This requires both a constexpr default construction *and* that the destructor // is permitted in constant evaluation, which implies a constexpr destructor. decltype( []{ constexpr T t{}; // ok only if T is literal *and* its dtor is constexpr (void)t; }(), void() ) > = true; consteval bool is_literal_type(std::meta::info type, std::meta::access_context ctx /* = unchecked */) { type = dealias(type); // Primary cases if (is_void_type(type)) return true; if (is_reference_type(type))return true; if (is_scalar_type(type)) return true; // Arrays: literal iff the element type is literal. if (is_array_type(type)) { return is_literal_type(remove_all_extents(type), ctx); } // Unions: aggregate union with a literal variant (or empty union) and non-volatile members. if (is_union_type(type)) { if (!is_aggregate_type(type)) return false; // matches the “aggregate union” bullet return union_has_literal_variant(type, ctx); } if (!class_members_are_literal(type, ctx)) { return false; // Destructor condition: C++20+ requires a constexpr destructor (trivial // so we implement a conservative check: // - accept trivially-destructible (always OK); // - otherwise, *attempt* to verify a constexpr destructor in the default-constructible case // via a substitution probe (__has_constexpr_dtor_if_default). if (!is_trivially_destructible_type(type)) { bool ok = false; // Try the default-constructible constexpr-dtor probe. // (If T isn't default-constructible but still has a constexpr dtor, we can't detect it.) ok = extract(substitute(^^__has_constexpr_dtor_if_default, { type })); if (!ok) return false; // conservative: fail if we cannot prove constexpr dtor } // Final disjunct for class types: // * aggregates are fine (no need to find a constexpr ctor explicitly), // * otherwise the standard allows “has at least one constexpr // or “lambda type”. P2996R13 doesn’t yet expose “is this constructor constexpr?” // nor “is this a lambda closure type?”, so be conservative and only accept aggregates. if (is_aggregate_type(type)) { return true; } // Without constructor-constexpr/lambda queries, we can't soundly accept // non-aggregate classes here. Be conservative. return false; } // Everything else (functions, etc.) — not literal. return false; consteval bool is_literal_class_type(std::meta::info t, std::meta::access_context ctx) { if (!(is_class_type(t) || is_union_type(t))) return false; return is_literal_type(t, ctx); consteval bool proposed::is_structural_type(std::meta::info t0) { // Ok to use unchecked access context, per LEWG Jan '26 telecon. constexpr auto ctx = std::meta::access_context::unchecked(); // 1) lvalue reference types are structural (top-level cv is ignored) if (is_lvalue_reference_type(t0)) { return true; } ``` `// Work with cv`‑`stripped, de-aliased type for the rest` `std::meta::info t = detail::canonicalize(t0);` ```cpp // 2) scalar types are structural if (is_scalar_type(t)) { return true; } ``` `// 3) literal class/union with the “public & non`‑`mutable &` `structural-or-array-of-structural” condition` `if` `(is_class_type(t) || is_union_type(t)) {` `if` `(!detail::is_literal_class_type(t, ctx)) {` `return` `false;` ```cpp } return detail::bases_ok(t) && detail::data_members_ok(t); } structural return false; } struct S { public: int x; }; struct Bad1 { private: int x; }; struct Bad2 { public: mutable int x; }; struct A { public: int a[2][3]; }; struct B : public S {}; // OK: public base of structural type struct C : private S {}; // not structural: private base static_assert( proposed::is_structural_type(^^int)); static_assert( proposed::is_structural_type(^^int&)); static_assert(!proposed::is_structural_type(^^int&&));// rvalue ref is not structural static_assert( proposed::is_structural_type(^^S)); static_assert(!proposed::is_structural_type(^^Bad1)); static_assert(!proposed::is_structural_type(^^Bad2)); static_assert( proposed::is_structural_type(^^A)); static_assert( proposed::is_structural_type(^^B)); static_assert(!proposed::is_structural_type(^^C)); int main() { } ``` [Godbolt link](https://godbolt.org/z/abGPKx1YM) * The implementation could be improved if metafunctions such as “is this a literal type”, “is this a lambda closure type” or “is this function constexpr” were available (*std::is_literal_type* was removed in C++20). * The proof of concept implementation ignores unsuitable inputs. If the input argument isn’t a reflection of a type, reflection metafunctions are supposed to throw an exception, as per the current specification. However, NB Comment US 129-191 proposes returning *false* instead. If the NB comment were accepted, the implementation of the new reflection metafunction would change only for the case when the input argument is the reflection of a non-type. The core logic for the happy case, when the input arguments are reflections of types, remains unchanged. ### 4. Type traits or metafunctions? [P2996](http://wg21.link/P2996) introduced several metafunctions *(consteval* functions that operate on *meta::info)* in ** that mirror existing type traits, such as *is_const_type(info), is_volatile_type(info),* and *is_trivially_copyable_type(info).* Type traits are typically implemented via templates (SFINAE) or compiler intrinsics, or a combination of both techniques. Metafunctions offer several benefits over traditional type traits when used to query the attributes of a type: * Simple and fixed interface that accepts a single, light-weight *meta::info* type parameter. * Avoid template instantiation and SFINAE-related memory bloat in most cases. * Provide a [wide range of building blocks](https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2025/p2996r13.html#stdmetainfo), as they can not only query types, but also values, and non-type entities like data members, functions, parameters, namespaces, etc. E.g., *std::meta::is_final(info type)* isn’t limited to “final class types”; it can also ask if a member function is final. Such a query is not expressible with type traits via a single trait. * Improved error handling by means of *meta::exception* that includes an error message, source and line number, thanks to [P3560](http://wg21.link/P3560). * Access control checks while querying type information, thanks to [P3547](http://wg21.link/p3547). Type queries could return different answers depending on the program point from where they’re evaluated. * Improved maintainability as the code is easier to read vs. SFINAE-based template waterfall logic. * Faster execution than a template-based implementation, though not as fast as using compiler intrinsics. Given that we now have two similar facilities in the standard library for inspecting type attributes, how should programmers approach them? Should type traits and reflection metafunctions evolve independently? Should one be implemented in terms of the other? More research and discussion needs to be carried out, though our preliminary thinking is that type traits and metafunctions should implement the abstract library API specification independently. They are different paradigms for metaprogramming, with different side effects. Metafunctions should not use library trait templates in their implementation, as that may confuse a programmer who expects to avoid template instantiations and encounter a different set of error logs. Conversely, type traits should not use reflection under the covers as that may surprise programmers who expect to see and handle the side effects of template metaprogramming. Add the new metafunctions to in [meta.reflection], in the list of unary type-property queries. 21.4 Reflection [meta.reflection] 21.4.1 Header synopsis [meta.syn] … ... + consteval bool is_structural_type(info type); ... Add a new metafunction is_structural_type to 21.4.17 Reflection type traits [meta.reflection.traits] 21.4.17 Reflection type traits[meta.reflection.traits] … // associated with [meta.unary.prop], type properties … consteval bool is_consteval_only_type(info type); + consteval bool is_structural_type(info type); consteval bool is_signed_type(info type); … Add is_structual type trait to [meta.type.synop] :::wording-add // [meta.unary.prop], type properties template struct is_const; .... + template struct is_structural; template struct is_signed; ... ::: Add *is_structural* type trait to [[meta.unary.prop]](https://eel.is/c++draft/meta.unary.prop) table 51 following paragraph 5: :::wording-add template constexpr bool is_const_v = is_const::value; template constexpr bool is_consteval_only_v = is_consteval_only::value; + template constexpr bool is_structural_v = is_structural::value; template ::: Add *is_structural* type trait to the table in [[meta.unary.prop]](https://eel.is/c++draft/meta.unary.prop#tab:meta.unary.prop-row) **Template** **Condition** **Preconditions** template struct is_consteval_only; T is consteval-only [([basic.types.general])](https://eel.is/c++draft/basic.types.general) remove_all_extents_t shall be a complete type or *cv* [void.](https://eel.is/c++draft/meta.unary.prop#tab:meta.unary.prop-row-11-column-3-sentence-1) :::wording-add + template + struct is_structural; ::: :::wording-add + T is a structural type ([temp.param]) ::: :::wording-add + remove_all_extents_t shall be a complete type or cv void. ::: struct is_signed; If is_arithmetic_v is true, the same result as T(-1) < T(0); otherwise, false ### 6. Feature-test Macros We propose adding a new feature-test macro to 17.3.2 [[version.syn](https://wg21.link/version.syn)] for the new type trait. ```cpp #define __cpp_lib_is_scoped_enum 202011L // + #define __cpp_lib_is_structural 2026XXL // #define __cpp_lib_is_sufficiently_aligned 202411L // freestanding, also in ``` Our experience in implementing new reflection metafunctions shows how reflection type queries can be implemented using other reflection metafunctions only (no intrinsics or SFINAE). It also shows the need for more “lower level” metafunctions not included in P2996, such as *is_constexpr(info),* *is_literal_type(info)* and *is_lambda_closure_type(info).* The approach of using reflection metafunctions to query type attributes appears promising, and likely to yield richer APIs than possible with prior techniques. We thank Barry Revzin, Andrei Zissu and LEWG for their review of the paper. [P2996R13] Barry Revzin, Wyatt Childers, Peter Dimov, Andrew Sutton, Faisal Vali, Daveed Vandevoorde, Dan Katz. 2025-01-13. Reflection for C++26. https://wg21.link/p2996r9 [P3547R1] Dan Katz, Ville Voutilainen. 2025-02-09. Modeling Access Control With Reflection. https://wg21.link/p3547r1 [P1061R10] Barry Revzin, Jonathan Wakely. 2024-11-24. Structured Bindings can introduce a Pack. https://wg21.link/p0144r2 [CppCon25Talk] Andrei Zissu. 2025-09. Reflection and Metaprogramming in Modern C++. https://www.youtube.com/watch?v=EK74rV1M7uc [N5028] 2025-10-2. SoV and Collated Comments - ISO_IEC CD 14882. https://wg21.link/N5028