Itanium C++ ABI


Contents


Acknowledgements


This document was originally developed jointly by an informal industry coalition consisting of (in alphabetical order) CodeSourcery, Compaq, EDG, HP, IBM, Intel, Red Hat, and SGI. Additional contributions were provided by a variety of individuals. It is now developed as an open-source project with contributions from a variety of individuals and companies.


Chapter 1: Introduction


In this document, we specify the Application Binary Interface (ABI) for C++ programs: that is, the object code interfaces between different user-provided C++ program fragments and between those fragments and the implementation-provided runtime and libraries. This includes the memory layout for C++ data objects, including both predefined and user-defined data types, as well as internal compiler generated objects such as virtual tables. It also includes function calling interfaces, exception handling interfaces, global naming, and various object code conventions.

In general, this document is meant to serve as a generic specification which can be used by C++ implementations on a variety of platforms. It does this by layering on top of a platform's base C ABI. However, it was originally written for the Itanium architecture, and some parts still directly make Itanium-specific or 64-bit-specific assumptions. There is an ongoing project to restate the entire C++ ABI specification in terms of portable C concepts that are defined in the C ABI. In the meantime, it is usually straightforward to recognize these unportable assumptions and translate them appropriately, e.g. by replacing a 64-bit pointer with a 32-bit pointer.

This document is not an authoritative definition of the C++ ABI for any particular platform. Platform vendors retain the ultimate power to define the C++ ABI for their platform. Platforms using this ABI for C++ should declare that they do so, either unmodified or with a certain set of changes.

While this ABI has generally stood up well, there are some parts of it that are now seen as mistakes. This document includes several recommendations for platforms adopting this ABI with no need to interoperate with existing C++ object code. These recommendations appear as follows:

Recommendation for new platforms: consider forbidding the use of function templates on your platform so that the ABI can remove these expression-mangling rules.

Platforms adopting any of these recommendations should describe the exact changes they've made in their platform ABI documentation, as the set of recommendations in this document may change over time.


1.1 Definitions

The descriptions below make use of the following definitions:

alignment of a type T (or object X)
A value A such that any object X of type T has an address satisfying the constraint that &X modulo A == 0.

base class of a class T
When this document refers to base classes of a class T, unless otherwise specified, it means T itself as well as all of the classes from which it is derived, directly or indirectly, virtually or non-virtually. We use the term  proper base class to exclude T itself from the list.

base object destructor of a class T
A function that runs the destructors for non-static data members of T and non-virtual direct base classes of T.

basic ABI properties of a type T
The basic representational properties of a type decided by the base C ABI, including its size, its alignment, its treatment by calling conventions, and the representation of pointers to it.

complete object destructor of a class T
A function that, in addition to the actions required of a base object destructor, runs the destructors for the virtual base classes of T.

deleting destructor of a class T
A function that, in addition to the actions required of a complete object destructor, calls the appropriate deallocation function (i.e,. operator delete) for T.

direct base class order
When the direct base classes of a class are viewed as an ordered set, the order assumed is the order declared, left-to-right.

diamond-shaped inheritance
A class has diamond-shaped inheritance iff it has a virtual base class that can be reached by distinct inheritance graph paths through more than one direct base.

dynamic class
A class requiring a virtual table pointer (because it or its bases have one or more virtual member functions or virtual base classes).

empty class
A class with no non-static data members other than empty data members, no unnamed bit-fields other than zero-width bit-fields, no virtual functions, no virtual base classes, and no non-empty non-virtual proper base classes.

empty data member
A potentially-overlapping non-static data member of empty class type.

inheritance graph
A graph with nodes representing a class and all of its subobjects, and arcs connecting each node with its direct bases.

inheritance graph order
The ordering on a class object and all its subobjects obtained by a depth-first traversal of its inheritance graph, from the most-derived class object to base objects, where:

Note that the traversal may be preorder or postorder. Unless otherwise specified, preorder (derived classes before their bases) is intended.

instantiation-dependent
An expression is instantiation-dependent if it is type-dependent or value-dependent, or it has a subexpression that is type-dependent or value-dependent. For example, if p is a type-dependent identifier, the expression sizeof(sizeof(p)) is neither type-dependent, nor value-dependent, but it is instantiation-dependent (and could turn out to be invalid if after substitution of template arguments p turns out to have an incomplete type). Similarly, a type expressed in source code is instantiation-dependent if the source form includes an instantiation-dependent expression. For example, the type form double[sizeof(sizeof(p))] (with p a type dependent identifier) is instantiation-dependent.

morally virtual
A subobject X is a morally virtual base of Y if X is either a virtual base of Y, or the direct or indirect base of a virtual base of Y.

nearly empty class
A class that contains a virtual pointer, but no other data except (possibly) virtual bases. In particular, it: Such classes may be primary base classes even if virtual, sharing a virtual pointer with the derived class.

non-trivial for the purposes of calls

A type is considered non-trivial for the purposes of calls if:

  • it has a non-trivial copy constructor, move constructor, or destructor, or
  • all of its copy and move constructors are deleted.

This definition, as applied to class types, is intended to be the complement of the definition in [class.temporary]p3 of types for which an extra temporary is allowed when passing or returning a type. A type which is trivial for the purposes of the ABI will be passed and returned according to the rules of the base C ABI, e.g. in registers; often this has the effect of performing a trivial copy of the type.

POD for the purpose of layout

In general, a type is considered a POD for the purposes of layout if it is a POD type (in the sense of ISO C++ [basic.types]). However, a type is not considered to be a POD for the purpose of layout if it is:

  • a POD-struct or POD-union (in the sense of ISO C++ [class]) with a bit-field whose declared width is wider than the declared type of the bit-field, or
  • an array type whose element type is not a POD for the purpose of layout, or
  • a POD-struct with one or more potentially-overlapping non-static data members.
Where references to the ISO C++ are made in this paragraph, the Technical Corrigendum 1 version of the standard is intended.

<b>NOTE</b>: There have been multiple published revisions to the ISO C++ standard, and each one has included a different definition of POD. To ensure interoperation of code compiled according to different revisions of the standard, it is necessary to settle on a single definition for a platform. A platform vendor may choose to follow a different revision of the standard, but by default, the definition of POD under this ABI is the definition from the 2003 revision (TC1).

Being tied to the TC1 definition of POD does not prevent compilers from being fully compliant with later revisions. This ABI uses the definition of POD only to decide whether to allocate objects in the tail-padding of a base-class subobject. While the standards have broadened the definition of POD over time, they have also forbidden the programmer from directly reading or writing the underlying bytes of a base-class subobject with, say, memcpy. Therefore, even in the most conservative interpretation, implementations may freely allocate objects in the tail padding of any class which would not have been POD in C++98. This ABI is in compliance with that.

potentially-overlapping subobject
A base class subobject or a non-static data member declared with the [[no_unique_address]] attribute.

primary base class
For a dynamic class, the unique base class (if any) with which it shares the virtual pointer at offset 0.

secondary virtual table
The instance of a virtual table for a base class that is embedded in the virtual table of a class derived from it.

templated entity
An entity that is defined or created within a template, such as:
  • an instantiation of a class, function, or variable template, including from a partial specialization, but not including an explicit specialization;
  • a member or friend function definition of a templated class;
  • an enumerator of a templated enum;
  • a local entity in a templated function; or
  • a lambda in a templated entity.

thunk
A segment of code associated (in this ABI) with a target function, which is called instead of the target function for the purpose of modifying parameters (e.g. this) or other parts of the environment before transferring control to the target function, and possibly making further modifications after its return. A thunk may contain as little as an instruction to be executed prior to falling through to an immediately following target function, or it may be a full function with its own stack frame that does a full call to the target function.

vague linkage
The treatment of entities -- e.g. inline functions, templates, virtual tables -- with external linkage that can be defined in multiple translation units, while the ODR requires that the program behave as if there were only a single definition.

virtual table (or vtable)
A dynamic class has an associated table (often several instances, but not one per object) which contains information about its dynamic attributes, e.g. virtual function pointers, virtual base class offsets, etc.

virtual table group
The primary virtual table for a class along with all of the associated secondary virtual tables for its proper base classes.


1.2 Limits

Various representations specified by this ABI impose limitations on conforming user programs. These include, for the 64-bit Itanium ABI:


1.3 Namespace and Header

This ABI specifies a number of type and function APIs supplemental to those required by the ISO C++ Standard. A header file named cxxabi.h will be provided by implementations that declares these APIs. The reference header file included with this ABI definition shall be the authoritative definition of the APIs.

These APIs will be placed in a namespace __cxxabiv1. The header file will also declare a namespace alias abi for __cxxabiv1. It is expected that users will use the alias, and the remainder of the ABI specification will use it as well.

In general, API objects defined as part of this ABI are assumed to be extern "C++". However, some (many?) are specified to be extern "C" if they:


1.4 Scope of This ABI

1.4.1 Runtime Libraries

The objective of a full ABI is to allow arbitrary mixing of object files produced by conforming implementations, by fully specifying the binary interface of application programs. We do not fully achieve this objective.

There are two principal reasons for this:

  1. We start from the Itanium processor-specific ABI as the standard for the underlying C interfaces. At this time, however, the psABI does not attempt to specify the supported C library interfaces.

  2. More fundamental is the definition of the Standard C++ Library. As the standard interface makes heavy use of templates, most user object files will end up with embedded template instantiations. Vendors are allowed to use helper functions and data in their implementations of these templates, and quite reasonably do so, with the result that a typical user object file will contain references to such helper objects specific to the implementation where compiled. We have not attempted to constrain the interface at this level, because we do not consider doing so feasible at this time.

Notwithstanding these problems, because this ABI does completely specify the data model and certain library interfaces that inherently interact between objects (e.g. construction, destruction, and exceptions), it is our intent that interoperation of object files produced by different compilers be possible in the following cases:

Even these cases can fail if the compiler makes use of implementation-defined library interfaces to implement runtime functionality without explicit user reference, e.g. a software divide function. We can distinguish between:

An implementation shall place its standard support library in a DSO named libcxa.so on Itanium systems, or in auxiliary DSOs automatically loaded by it. It shall place implicit compiler support in a library separate from the standard support library, with any external names chosen to avoid conflicts between vendors (e.g. by including a vendor identifier as part of the names). This allows a program to function properly if linked with the target's standard support library and the implicit compiler support libraries from any implementations used to build components.

1.4.2 Export Templates

This ABI does not specify the treatment of export templates, as there are no working implementations to serve as models at this time. We hope to address this weakness in the future when implementation experience is available.


1.5 Base Documents

A number of other documents provide a basis on which this ABI is built, and are occasionally referenced herein:


Chapter 2: Data Layout


2.1 General

In what follows, we define the memory layout for C++ data objects. Specifically, for each type, we specify the following information about an object O of that type:

For purposes internal to the specification, we also specify:


2.2 POD Data Types

The size and alignment of a type which is a POD for the purpose of layout is as specified by the base C ABI, with the following provisos:

The dsize, nvsize, and nvalign of these types are defined to be their ordinary size and alignment. These properties only matter for non-empty class types that are used as base classes. We ignore tail padding for PODs because the standard before the resolution of CWG issue 43 did not allow us to use it for anything else and because it sometimes permits faster copying of the type.


2.3 Member Pointers

2.3.1 Data Member Pointers

The basic ABI properties of data member pointer types are those of ptrdiff_t.

A data member pointer is represented as the data member's offset in bytes from the address point of an object of the base type, as a ptrdiff_t.

A null data member pointer is represented as an offset of -1.

Note that by [dcl.init], "zero initialization" of a data member pointer object stores a null pointer value into it. Under this representation, that value has a non-zero bit representation. On most modern platforms, data member pointers are the only type with this property.

Base-to-derived and derived-to-base conversions of a non-null data member pointer can be performed by adding or subtracting (respectively) the static offset of the base within the derived class. The C++ standard does not permit base-to-derived and derived-to-base conversions of member pointers to cross a virtual base relationship, and so a static offset is always known.

Data member pointers that identify members of their class will always store non-negative offsets. Unfortunately, it is possible to apply conversions to a non-null data member pointer that will cause it to hold a negative offset. If this value is -1, the member pointer will subsequently be treated as a null pointer. This is considered an irreparable defect in this ABI. Recommendation for new platforms: consider using a different representation for data member pointers, such as left-shifting the offset by one and using a non-zero low bit to indicate a non-null value.

It is relatively difficult to demonstrate this defect in well-defined code. It is possible to convert a member pointer to a derived class and then cast it back it to a different base class; if the second base class is stored after the first, the resulting offset will be negative. However, this cast has undefined behavior because the member is no longer a member of a base or derived class of the member pointer's class. To demonstrate the defect, either an empty base class or an empty data member must be involved. For example:

    struct alignas(2) B1 {};

    struct B2 : B1 {};
    struct B3 : B1 {};

    struct D : B2, B3 {
        char a, b;
    };

    // The offset in D of the B3 base subobject is 2, but the
    // offset of the data member b is 1.
    auto mptr = static_cast<char B3::*>(&D::b);

2.3.2 Member Function Pointers

Several different representions of member function pointers are in use. The standard representation relies on several assumptions about the platform, such as that the low bit of a function pointer to a non-virtual member function is always zero. For platforms where this is not reasonable to guarantee, an alternate representation must be used. One such representation, used on the 32-bit ARM architecture, is also described here.

In all representations, the basic ABI properties of member function pointer types are those of the following class, where fnptr_t is the appropriate function-pointer type for a member function of this type:

  struct {
    fnptr_t ptr;
    ptrdiff_t adj;
  };

A member function pointer for a non-virtual member function is represented with ptr set to a pointer to the function, using the base ABI's representation of function pointers.

In the standard representation, a member function pointer for a virtual function is represented with ptr set to 1 plus the function's v-table entry offset (in bytes), converted to a function pointer as if by reinterpret_cast<fnptr_t>(uintfnptr_t(1 + offset)), where uintfnptr_t is an unsigned integer of the same size as fnptr_t.

In both of these cases, adj stores the offset (in bytes) which must be added to the this pointer before the call.

In the standard representation, a null member function pointer is represented with ptr set to a null pointer. The value of adj is unspecified for null member function pointers.

The standard representation relies on some assumptions which are true for most platforms:

However, there are exceptions. For example, on the 32-bit ARM architecture, the low bit of a function pointer determines whether the function begins in THUMB mode. Such platforms must use an alternate representation.

In the 32-bit ARM representation, the this-adjustment stored in adj is left-shifted by one, and the low bit of adj indicates whether ptr is a function pointer (including null) or the offset of a v-table entry. A virtual member function pointer sets ptr to the v-table entry offset as if by reinterpret_cast<fnptr_t>(uintfnptr_t(offset)). A null member function pointer sets ptr to a null function pointer and must ensure that the low bit of adj is clear; the upper bits of adj remain unspecified.

A member function pointer is null if ptr is equal to a null function pointer and (only when using the 32-bit ARM representation) the low bit of adj is clear.

Two member function pointers are equal if they are both null or if their corresponding values of ptr and adj are equal. Note that the C++ standard does not require member pointers to the same virtual member function to compare equal; implementations using this ABI will do so, but only if the member pointers are built using the same v-table offset, which they may not be in the presence of multiple inheritance or overrides with covariant return types.

Base-to-derived and derived-to-base conversions of a member function pointer can be performed by adding or subtracting (respectively) the static offset of the base within the derived class to the stored this-adjustment value. In the standard representation, this simply means adding it to adj; in the 32-bit ARM representation, the addend must be left-shifted by one. Because the adjustment does not factor into whether a member function pointer is null, this addition can be done unconditionally when performing a conversion.

A call is performed as follows:

  1. Add the stored adjustment to the this address.
  2. If the member pointer stores a v-table entry offset, load the v-table from the adjusted this address and call the v-table entry at the stored offset.
  3. Otherwise, call the stored function pointer.


2.4 Non-POD Class Types

For a class type C which is not a POD for the purpose of layout, assume that all component types (i.e. proper base classes and non-static data member types) have been laid out, defining size, data size, non-virtual size, alignment, and non-virtual alignment. (See the description of these terms in General above.) Layout (of type C) is done using the following procedure.

  1. Initialization

    1. Initialize sizeof(C) to zero, align(C) to one, dsize(C) to zero.

    2. If C is a dynamic class type:

      1. Identify all virtual base classes, direct or indirect, that are primary base classes for some other direct or indirect base class. Call these indirect primary base classes.

      2. If C has a dynamic base class, attempt to choose a primary base class B. It is the first (in direct base class order) non-virtual dynamic base class, if one exists. Otherwise, it is a nearly empty virtual base class, the first one in (preorder) inheritance graph order which is not an indirect primary base class if any exist, or just the first one if they are all indirect primaries.

      3. If C has no primary base class, allocate the virtual table pointer for C at offset zero, and set sizeof(C), align(C), and dsize(C) to the appropriate values for a pointer (all 8 bytes for Itanium 64-bit ABI).

    <b>NOTE</b>: Case (2b) above is now considered to be an error in the design. The use of the first indirect primary base class as the derived class' primary base does not save any space in the object, and will cause some duplication of virtual function pointers in the additional copy of the base classes virtual table.

    The benefit is that using the derived class virtual pointer as the base class virtual pointer will often save a load, and no adjustment to the this pointer will be required for calls to its virtual functions.

    It was thought that 2b would allow the compiler to avoid adjusting this in some cases, but this was incorrect, as the virtual function call algorithm requires that the function be looked up through a pointer to a class that defines the function, not one that just inherits it. Removing that requirement would not be a good idea, as there would then no longer be a way to emit all thunks with the functions they jump to. For instance, consider this example:

    struct A { virtual void f(); };
    struct B : virtual public A { int i; };
    struct C : virtual public A { int j; };
    struct D : public B, public C {};
    

    When B and C are declared, A is a primary base in each case, so although vcall offsets are allocated in the A-in-B and A-in-C vtables, no this adjustment is required and no thunk is generated. However, inside D objects, A is no longer a primary base of C, so if we allowed calls to C::f() to use the copy of A's vtable in the C subobject, we would need to adjust this from C* to B::A*, which would require a third-party thunk. Since we require that a call to C::f() first convert to A*, C-in-D's copy of A's vtable is never referenced, so this is not necessary.

  2. Allocation of Members Other Than Virtual Bases

    For each data component D (first the primary base of C, if any, then the non-primary, non-virtual direct base classes in declaration order, then the non-static data members and unnamed bit-fields in declaration order), allocate as follows:

    1. If D is a (possibly unnamed) bit-field whose declared type is T and whose declared width is n bits:

      There are two cases depending on sizeof(T) and n:

      1. If sizeof(T)*8 >= n, the bit-field is allocated as required by the base C ABI, subject to the constraint that a bit-field is never placed in the tail padding of a base class of C.

        If dsize(C) > 0, and the byte at offset dsize(C) - 1 is partially filled by a bit-field, and that bit-field is also a data member declared in C (but not in one of C's proper base classes), the next available bits are the unfilled bits at offset dsize(C) - 1. Otherwise, the next available bits are at offset dsize(C).

        Update align(C) to max (align(C), align(T)).

      2. If sizeof(T)*8 < n, let T' be the largest integral POD type with sizeof(T')*8 <= n. The bit-field is allocated starting at the next offset aligned appropriately for T', with length n bits. The first sizeof(T)*8 bits are used to hold the value of the bit-field, followed by n - sizeof(T)*8 bits of padding.

        Update align(C) to max (align(C), align(T')).

      In either case, update dsize(C) to include the last byte containing (part of) the bit-field, and update sizeof(C) to max(sizeof(C),dsize(C)).

    2. If D is not an empty base class and D is not an empty data member:

      Start at offset dsize(C), incremented if necessary for alignment to nvalign(D) for base classes or to align(D) for data members. Place D at this offset unless doing so would result in two components (direct or indirect) of the same type having the same offset. If such a component type conflict occurs, increment the candidate offset by nvalign(D) for base classes or by align(D) for data members and try again, repeating until success occurs (which will occur no later than sizeof(C) rounded up to the required alignment).

      If D is a base class, this step allocates only its non-virtual part, i.e. excluding any direct or indirect virtual bases.

      If D is a base class, update sizeof(C) to max (sizeof(C), offset(D)+nvsize(D)). Otherwise, if D is a potentially-overlapping data member, update sizeof(C) to max (sizeof(C), offset(D)+max (nvsize(D), dsize(D))). Otherwise, if D is a data member, update sizeof(C) to max (sizeof(C), offset(D)+sizeof(D)).

      If D is a base class (not empty in this case), update dsize(C) to offset(D)+nvsize(D), and align(C) to max (align(C), nvalign(D)). If D is a potentially-overlapping data member, update dsize(C) to offset(D)+max (nvsize(D), dsize(D)), align(C) to max (align(C), align(D)). If D is any other data member, update dsize(C) to offset(D)+sizeof(D), align(C) to max (align(C), align(D)).

    3. If D is an empty proper base class or an empty data member:

      Its allocation is similar to case (2) above, except that additional candidate offsets are considered before starting at dsize(C). First, attempt to place D at offset zero. If unsuccessful (due to a component type conflict), proceed with attempts at dsize(C) as for non-empty bases and members. As for that case, if there is a type conflict at dsize(C) (with alignment updated as necessary), increment the candidate offset by nvalign(D), and try again, repeating until success occurs.

      Once offset(D) has been chosen, update sizeof(C) to max (sizeof(C), offset(D)+sizeof(D)), and update align(C) to max (alignof(C), nvalign(D)) for a base class or max (alignof(C), align(D)) for a data member. Since D is empty, no update of dsize(C) is needed.

    After all such components have been allocated, set nvalign(C) = align(C) and nvsize(C) = sizeof(C). The values of nvalign(C) and nvsize(C) will not change during virtual base allocation. Note that nvsize(C) need not be a multiple of nvalign(C).

  3. Virtual Base Allocation

    Finally allocate any direct or indirect virtual base classes (except the primary base class or any indirect primary base classes) as we did non-virtual base classes in step II-2 (if not empty) or II-3 (if empty), in inheritance graph order. Update sizeof(C) to max (sizeof(C), offset(D)+nvsize(D)). If non-empty, also update align(C) and dsize(C) as in II-2.

    The primary base class has already been allocated in I-2b. Any indirect primary base class E of the current class C, i.e. one that has been chosen as the primary base class of some other base class (direct or indirect, virtual or non-virtual) of C, will be allocated as part of that other base class, and is not allocated here. If E is a primary base class of more than one other base, the instance used as its allocation in C shall be the first such in the inheritance graph order.

    Consider:

    
      struct R { virtual void r (); };
      struct S { virtual void s (); };
      struct T : virtual public S { virtual void t (); };
      struct U : public R, virtual public T { virtual void u (); };
    
    
    R is the primary base class for U since it is the first direct non-virtual dynamic base. Then, since an inheritance-order walk of U is { U, R, T, S } the T base is allocated next. Since S is a primary base of T, there is no need to allocate it separately. However, given:
    
      struct V : public R, virtual public S, virtual public T {
        virtual void v ();
      };
    
    
    the inheritance-order walk of V is { V, R, S, T }. Nevertheless, although S is considered for allocation first as a virtual base, it is not allocated separately because it is a primary base of T, another base. Thus sizeof (V) == sizeof (U), and the full layout is equivalent to the C struct:
    
      struct X {
        R r;
        T t;
      };
    
    

  4. Finalization

    For each potentially-overlapping non-static data member D of C, update sizeof(C) to max (sizeof(C), offset(D)+sizeof(D)). Example:

    
      struct alignas(16) A { ~A(); }; // dsize 0, nvsize 0, size 16
      struct B : A {}; // dsize 0, nvsize 16, size 16
      struct X : virtual A, virtual B {}; // dsize 8, nvsize 8, size 32
      struct Y { [[no_unique_address]] X x; char c; }; // dsize 9, nvsize 9, size 32
    

    Then, round sizeof(C) up to a non-zero multiple of align(C). If C is a POD, but not a POD for the purpose of layout, set dsize(C) = nvsize(C) = sizeof(C).


2.5 Virtual Table Layout

2.5.1 General

A virtual table (vtable) is a table of information used to dispatch virtual functions, to access virtual base class subobjects, and to access information for runtime type identification (RTTI). Each class that has virtual member functions or virtual bases has an associated set of virtual tables. There may be multiple virtual tables for a particular class, if it is used as a base class for other classes. However, the virtual table pointers within all the objects (instances) of a particular most-derived class point to the same set of virtual tables.

A virtual table consists of a sequence of offsets, data pointers, and function pointers, as well as structures composed of such items. We will describe below the sequence of such items. Their offsets within the virtual table are determined by that allocation sequence and the natural ABI size and alignment, just as a data struct would be. In particular:

In general, what we consider the address of a virtual table (i.e. the address contained in objects pointing to a virtual table) may not be the beginning of the virtual table. We call it the address point of the virtual table. The virtual table may therefore contain components at either positive or negative offsets from its address point.

2.5.2 Virtual Table Components and Order

This section describes the usage and relative order of various components that may appear in virtual tables. Precisely which components are present in various possible virtual tables is specified in the next section. If present, components are present in the order described, except for the exceptions specified.

Following the primary virtual table of a derived class are secondary virtual tables for each of its proper base classes, except any primary base(s) with which it shares its primary virtual table. These are copies of the virtual tables for the respective base classes (copies in the sense that they have the same layout, though the fields may have different values). We call the collection consisting of a primary virtual table along with all of its secondary virtual tables a virtual table group. The order in which they occur is the same as the order in which the base class subobjects are considered for allocation in the derived object:

This ABI does not make guarantees about the layout of other virtual tables in a virtual table group relative to a virtual table pointer in an object or a VTT. It guarantees only the layout of the global symbol for that virtual table group. It does not guarantee that the virtual table pointers actually installed in an object or a VTT will point into that global symbol.

2.5.3 Virtual Table Construction

In this section, we describe how to construct the virtual table for an class, given virtual tables for all of its proper base classes. To do so, we divide classes into several categories, based on their base class structure.

Category 0: Trivial
Structure:

Such a class has no associated virtual table, and an object of such a class contains no virtual pointer.

Category 1: Leaf
Structure:

The virtual table contains offset-to-top and RTTI fields followed by virtual function pointers (as specified above).

Category 2: Non-Virtual Bases Only
Structure:

The class has a virtual table for each proper base class that has a virtual table. The secondary virtual table for a base class B has the same contents as the primary virtual table for B, except that:

For a proper base class Base, and a derived class Derived for which we are constructing this set of virtual tables, we shall refer to the virtual table for Base as Base-in-Derived. The virtual pointer of each base subobject of an object of the derived class will point to the corresponding base virtual table in this set.

The primary virtual table for the derived class contains entries for each of the functions in the primary base class virtual table, replaced by new overriding functions as appropriate. Following these entries, there is an entry for each virtual function pointer for the derived class (as specified above).

<b>NOTE</b>: The primary virtual table can be viewed as two virtual tables accessed from a shared virtual table pointer.

<b>NOTE</b>: A benefit of replicated virtual function entries (i.e., entries that appear both in the primary virtual table and in a secondary virtual table) is that they reduce the number of this pointer adjustments during virtual calls. Without replication, there would be more cases where the this pointer would have to be adjusted to access a secondary virtual table prior to the call. These additional cases would be exactly those where the function is overridden in the derived class, implying an additional thunk adjustment back to the original pointer. Replication saves two 'this' adjustments for each virtual call to an overridden function originally introduced by a non-primary proper base class.

Category 3: Virtual Bases Only

Structure:

The class has a virtual table for each virtual base class that has a virtual table. These are all secondary virtual tables, because there are no empty or nearly empty base classes to be primary, and they are constructed from copies of the base class full object virtual tables according to the same rules as in Category 2, except that the virtual table for a virtual base A also includes a vcall offset entry for each virtual function represented in A's primary virtual table and the secondary virtual tables from A's non-virtual bases.

The vcall offsets in the secondary virtual table for a virtual base A are ordered as described next. We describe the ordering from the entry closest to the virtual table address point to that furthest. Since the vcall offsets precede the virtual table address point, this means that the memory address order is the reverse of that described.

If the above listing of vcall offsets includes more than one for a particular virtual function signature, only the first one (closest to the virtual table address point) is allocated. That is, an offset from primary base P (and its non-virtual bases) eliminates any from A or its other bases, an offset from A eliminates any from the non-primary bases, and an offset from a non-primary base B of A eliminates any from the bases of B.

Note that there are no vcall offsets for virtual functions declared in a virtual base class V of A and never overridden within A or its non-virtual bases. Calls to such functions will use the vcall offset in V's virtual table.

The class also has a virtual table that is not copied from the virtual base class virtual tables. This virtual table is the primary virtual table of the class and is addressed by the virtual table pointer at the top of the object, which is not shared because there are no nearly empty virtual bases to be primary. It holds the virtual function pointer entries for the class. This includes all entries overridden from base classes, because there is no primary base class; such entries are known as replicated entries because they are already in the secondary virtual tables of the class.

The primary virtual table also has virtual base offset entries to allow finding the virtual base subobjects. There is one virtual base offset entry for each virtual base class, direct or indirect. The entries are in the reverse of the inheritance graph order. That is, the entry for the leftmost virtual base is closest to the address point of the virtual table.

Category 4: Complex

Structure:

The rules for constructing virtual tables of the class are a combination of the rules from Categories 2 and 3, and can generally be determined inductively. The differences are mostly due to the fact that virtual base classes can now have (nearly empty) primary bases:

<b>NOTE</b>: For an S-as-T virtual table, the vbase offset entries from the primary virtual table for T are replaced with appropriate offsets given the completed hierarchy.

<b>NOTE</b>: Consider the following inheritance hierarchy:

  struct S { virtual void f() };
  struct T : virtual public S {};
  struct U : virtual public T {};
  struct V : public T, virtual public U {};

T's virtual table contains a virtual base offset for S. U's virtual table contains virtual base offsets for S and T. V's virtual table contains virtual base offsets for S, U, and T (in reverse inheritance graph preorder), where the vbase offset for T is for the virtual base of U, not for the non-virtual direct base of V.

Consider in addition:

  struct W : public T {};

T is a primary base class for W. Therefore, its virtual base offset for S in its embedded T-in-W virtual table is the only one present.

<b>NOTE</b>: The above-described virtual table group layout would allow all non-virtual secondary base class virtual tables in a group to be accessed from a virtual pointer for one of them, since the relative offsets would be fixed. (Since the primary virtual table could end up being embedded, as the primary base class virtual table, in another virtual table with additional virtual pointers separating it from its secondary virtual tables, this observation is not true of the primary virtual table.) However, since construction virtual table groups may be organized differently (see below), an implementation may not depend on this relationship between secondary virtual tables. This tradeoff was made because the space savings resulting from not requiring construction virtual tables to occur in complete groups was considered more important than potential sharing of virtual pointers.


2.6 Virtual tables During Object Construction

2.6.1 General

In some situations, a special virtual table called a construction virtual table is used during the execution of proper base class constructors and destructors. These virtual tables are for specific cases of virtual inheritance.

During the construction of a class object, the object assumes the type of each of its proper base classes, as each base class subobject is constructed. RTTI queries in the base class constructor will return the type of the base class, and virtual calls will resolve to member functions of the base class rather than the complete class. RTTI queries, dynamic casts and virtual calls of the object under construction statically converted to bases of the base under construction will dynamically resolve to the type of the base under construction. Normally, this behavior is accomplished by setting, in the base class constructor, the object's virtual table pointers to the addresses of the virtual tables for the base class.

However, if the base class has direct or indirect virtual bases, the virtual table pointers have to be set to the addresses of construction virtual tables. This is because the normal proper base class virtual tables may not hold the correct virtual base index values to access the virtual bases of the object under construction, and adjustment addressed by these virtual tables may hold the wrong this parameter adjustment if the adjustment is to cast from a virtual base to another part of the object. The problem is that a complete object of a proper base class and a complete object of a derived class do not have virtual bases at the same offsets.

A construction virtual table holds the virtual function addresses, offset-to-top, and RTTI information associated with the base class, and virtual base offsets and addresses of adjustor entry points with their parameter adjustments associated with objects of the complete class.

To ensure that the virtual table pointers are set to the appropriate virtual tables during proper base class construction, a table of virtual table pointers, called the VTT, which holds the addresses of construction and non-construction virtual tables is generated for the complete class. The constructor for the complete class passes to each proper base class constructor a pointer to the appropriate place in the VTT where the proper base class constructor can find its set of virtual tables. Construction virtual tables are used in a similar way during the execution of proper base class destructors.

NOTE When a complete object constructor is constructing a virtual base, it must be wary of using the vbase offsets in the virtual table, since the possibly shared virtual pointer may point to a construction virtual table of an unrelated base class. For instance, in


  struct S {};
  struct T: virtual S {};
  struct U {};
  struct V: virtual T, virtual U {};
the virtual pointers for T and V are in the same place. When V's constructor is about to construct U, that virtual pointer points to a virtual table for T, and therefore cannot be used to locate U.

2.6.2 VTT Order

An array of virtual table addresses, called the VTT, is declared for each class type that has indirect or direct virtual base classes. (Otherwise, each proper base class may be initialized using its complete object virtual table group.)

The elements of the VTT array for a class D are in this order:

  1. Primary virtual pointer: address of the primary virtual table for the complete object D.

  2. Secondary VTTs: for each direct non-virtual proper base class B of D that requires a VTT, in declaration order, a sub-VTT for B-in-D, structured like the main VTT for B, with a primary virtual pointer, secondary VTTs, and secondary virtual pointers, but without virtual VTTs.

    <b>NOTE</b>: This construction is applied recursively.

  3. Secondary virtual pointers: for each base class X which (a) has virtual bases or is reachable along a virtual path from D, and (b) is not a non-virtual primary base, the address of the virtual table for X-in-D or an appropriate construction virtual table.

    X is reachable along a virtual path from D if there exists a path X, B1, B2, ..., BN, D in the inheritance graph such that at least one of X, B1, B2, ..., or BN is a virtual base class.

    The order in which the virtual pointers appear in the VTT is inheritance graph preorder.

    <b>NOTE</b>: There are virtual pointers for direct and indirect base classes. Although primary non-virtual bases do not get secondary virtual pointers, they do not otherwise affect the ordering.

    Primary virtual bases require a secondary virtual pointer in the VTT because the derived class with which they will share a virtual pointer is determined by the most derived class in the hierarchy.

    Secondary virtual pointers may be required for base classes that do not require secondary VTTs. A virtual base with no virtual bases of its own does not require a VTT, but does require a virtual pointer entry in the VTT.

  4. Virtual VTTs: For each proper virtual base classes in inheritance graph preorder, construct a sub-VTT as in (2) above.

    <b>NOTE</b>: The virtual VTT addresses come last because they are only passed to the virtual base class constructors for the complete object.

Each virtual table address in the VTT is the address to be assigned to the respective virtual pointer, i.e. the address past the end of the typeinfo pointer (the address of the first virtual function pointer, if there are any), not of the first vcall offset.

<b>NOTE</b>: It is required that the VTT for a complete class D be identical in structure to the sub-VTT for the same class D as a base of another class E derived from it, so that the constructors for D can depend on that structure. Therefore, the various components of its VTT are present based on the rules given, even if they point to the D complete object virtual table or its secondary virtual tables. That is, secondary VTTs are present for all bases with virtual bases (including the virtual bases themselves, which have their secondary VTTs in the virtual VTT section), and secondary virtual pointers are present for all bases with either virtual bases or virtual function declarations overridden along a virtual path. The only exception is that a primary non-virtual base class does not require a secondary virtual pointer.

Parts (1) and (3) of a primary (not secondary, i.e. nested) VTT, that is the primary and secondary virtual pointers, are used for the final initialization of an object's virtual pointers before the full-object initialization and later use, and must therefore point to the main virtual table group for the class. Those bases which do not have secondary virtual pointers in the VTT have their virtual pointers explicitly initialized to the main virtual table group by the constructors (see Subobject Construction and Destruction).

The virtual pointers in the secondary VTTs and virtual VTTs are used for subobject construction, and may always point to special construction virtual tables laid out as described in the following subsections. However, it will sometimes be possible to use either the full-object virtual table for the base class, or its secondary virtual table for the full class being constructed. This ABI does not specify a choice, nor does it specify names for the construction virtual tables, so the constructors must use the VTT rather than assuming that a particular construction virtual table exists.

For example, suppose we have the following hierarchy:

  class A1 { int i; };
  class A2 { int i; virtual void f(); };
  class V1 : public A1, public A2 { int i; };
	// A2 is primary base of V1, A1 is non-polymorphic
  class B1 { int i; };
  class B2 { int i; };
  class V2 : public B1, public B2, public virtual V1 { int i; };
	// V2 has no primary base, V1 is secondary base
  class V3 {virtual void g(); };
  class C1 : public virtual V1 { int i; };
	// C1 has no primary base, V1 is secondary base
  class C2 : public virtual V3, public virtual V2 { int i; };
	// C2 has V3 primary (nearly-empty virtual) base, V2 is secondary base
  class X1 { int i; };
  class C3 : public X1 { int i; };
  class D : public C1, public C2, public C3 { int i;  };
	// C1 is primary base, C2 is secondary base, C3 is non-polymorphic

Then the VTT for D would appear in the following order, where indenting indicates the sub-VTT structure, and asterisks (*) indicate that construction virtual tables instead of complete object virtual tables are required.
  // 1. Primary virtual pointer:
  [0] D has virtual bases (complete object vptr)

  // 2. Secondary VTTs:
  [1]  C1 * (has virtual base)
  [2]     V1-in-C1 in D (secondary vptr)

  [3]  C2 * (has virtual bases)
  [4]    V3-in-C2 in D (primary vptr)
  [5]    V2-in-C2 in D (secondary vptr)
  [6]    V1-in-C2 in D (secondary vptr)

  // 3. Secondary virtual pointers:
    // (no C1-in-D -- primary base)
  [7]    V1-in-D (V1 is virtual)
  [8]  C2-in-D (preorder; has virtual bases)
  [9]    V3-in-D (V3 is virtual)
  [10]    V2-in-D (V2 is virtual)
    // (For complete object D VTT, these all can point to the
    // secondary vtables in the D vtable, the V3-in-D entry
    // will be the same as the C2-in-D entry, as that is the active
    // V3 virtual base in the complete object D.  In the sub-VTT for
    // D in a class derived from D, some might be construction
    // virtual tables.)

  // 4. Virtual VTTs:
    // (V1 has no virtual bases).
  [11] V2 * (V2 has virtual bases)
  [12]   V1-in-V2 in D * (secondary vptr, V1 is virtual)
	   (A2 is primary base of V1)
    // (V3 has no virtual bases)

If A2 is a virtual base of V1, the VTT will contain more elements (exercise left to the astute reader).

2.6.3 Construction Virtual Table Layout

The construction virtual tables for a complete object are emitted in the same object file as the virtual table. So the virtual table structures for a complete object of class C include, in no particular order:

The VTT array is referenced via its own mangled external name, and the construction virtual tables are accessed via the VTT array, so the latter do not have external names.

2.6.4 Construction Virtual Table entries

The construction virtual table group for a proper base class subobject B (of derived class D) does not have the same entries in the same order as the main virtual table group for a complete object B, as described in Virtual Table Layout above. Some of the base class subobjects may not need construction virtual tables, which will therefore not be present in the construction virtual table group, even though the subobject virtual tables are present in the main virtual table group for the complete object.

The values of some construction virtual table entries will differ from the corresponding entries in either the main virtual table group for B or the virtual table group for B-in-D, primarily because the virtual bases of B will be at different relative offsets in a D object than in a standalone B object, as follows:

  1. Virtual base class offsets reflect the positions of the virtual base classes in the full D object.
  2. Similarly, vcall offsets reflect the relative positions of the overridden and overriding classes within the complete object D.
  3. The offset-to-top fields refer to B (and B's in particular will therefore be zero).
  4. The RTTI pointers point to B's RTTI.
  5. Only functions in B and its base classes are considered for virtual function resolution.


2.7 Array Operator new Cookies

When operator new is used to create a new array, a cookie is usually stored to remember the allocated length (number of array elements) so that it can be deallocated correctly.

Specifically:

Given the above, the following is pseudocode for processing new(ARGS) T[n]:

  if T has a trivial destructor (C++ standard, 12.4/3)
    padding = 0
  else if we're using ::operator new[](size_t, void*)
    padding = 0
  else
    padding = max(sizeof(size_t), alignof(T))

  p = operator new[](n * sizeof(T) + padding, ARGS)
  p1 = (T*) ( (char *)p + padding )

  if padding > 0
    *( (size_t *)p1 - 1) = n

  for i = [0, n)
    create a T, using the default constructor, at p1[i]

  return p1


2.8 Initialization Guard Variables

If a function-scope static variable or a static data member with vague linkage (i.e., a static data member of a class template) is dynamically initialized, then there is an associated guard variable which is used to guarantee that construction occurs only once. The guard variable's name is mangled based on the mangling of the guarded object name. Thus, for function-scope static variables, if multiple instances of the function body are emitted (e.g., due to inlining), each function uses the same guard variable to ensure that the function-scope static is initialized only once. Similarly, if a static data member is instantiated in multiple object files, the initialization code in each object file will use the same guard variable to ensure that the static data member is initialized only once.

The size of the guard variable is 64 bits. The first byte (i.e. the byte at the address of the full variable) shall contain the value 0 prior to initialization of the associated variable, and 1 after initialization is complete. Usage of the other bytes of the guard variable is implementation-defined.

See Section 3.3.3 for the API for references to this guard variable.


2.9 Run-Time Type Information (RTTI)

2.9.1 General

The C++ programming language definition implies that information about types be available at run time for three distinct purposes:

  1. to support the typeid operator,
  2. to match an exception handler with a thrown object, and
  3. to implement the dynamic_cast operator.
(c) only requires type information about dynamic class types, but (a) and (b) may apply to other types as well; for example, when a pointer to an int is thrown, it can be caught by a handler that catches "int const*".

It is intended that two type_info pointers point to equivalent type descriptions if and only if the pointers are equal. An implementation must satisfy this constraint, e.g. by using symbol preemption, COMDAT sections, or other mechanisms.

<b>NOTE</b>: Note that the full structure described by an RTTI descriptor may include incomplete types not required by the Standard to be completed, although not in contexts where it would cause ambiguity. Therefore, any cross-references within the RTTI to types not known to be complete must be weak symbol references.

2.9.2 Place of Emission

It is desirable to minimize the number of places where a particular bit of RTTI is emitted. For dynamic class types, a similar problem occurs for virtual function tables, and hence the RTTI descriptor should be emitted with the primary virtual table for that type. For other types, they must be emitted at the location where their use is implied: the object file containing the typeid, throw or catch.

Basic type information (e.g. for "int", "bool", etc.) will be kept in the run-time support library. Specifically, the run-time support library should contain type_info objects for the types X, X* and X const*, for every X in: void, std::nullptr_t, bool, wchar_t, char, unsigned char, signed char, short, unsigned short, int, unsigned int, long, unsigned long, long long, unsigned long long, float, double, long double, char8_t, char16_t, char32_t, and the IEEE 754r decimal and half-precision floating point types. Each of the type_info objects for X shall have type abi::__fundamental_type_info (or a type derived therefrom), whereas the objects corresponding to X* and X const* shall have type abi::__pointer_type_info (or a type derived therefrom). (Note that various other type_info objects for class types may reside in the run-time support library by virtue of the preceding rules, e.g. that of std::bad_alloc.)

2.9.3 The typeid Operator

The typeid operator produces a reference to a std::type_info structure with the following public interface (18.5.1):


  namespace std {
    class type_info {
      public:
	virtual ~type_info();
	bool operator==(const type_info &) const;
	bool operator!=(const type_info &) const;
	bool before(const type_info &) const;
	const char* name() const;
      private:
	type_info (const type_info& rhs);
	type_info& operator= (const type_info& rhs);
    };
  }

After linking and loading, only one std::type_info structure is accessible via the external name defined by this ABI for any particular complete type symbol (see Vague Linkage). Therefore, except for direct or indirect pointers to incomplete types, the equality and inequality operators can be written as address comparisons when operating on those type_info objects: two type_info structures describe the same type if and only if they are the same structure (at the same address). However, in the case of pointer types, directly or indirectly pointing to incomplete class types, a more complex comparison is required, described below with the RTTI layout of pointer types.

The name() member function returns the address of an NTBS, unique to the type, containing the mangled name of the type. The mangled name of the NTBS is also defined by the ABI to allow consistent reference to it, and the Vague Linkage section specifies how to produce a unique copy.

In a flat address space (such as that of the Itanium architecture), the operator==, operator!=, and before() members are easily implemented in terms of an address comparison of the name NTBS.

This implies that the type information must keep a description of the public, unambiguous inheritance relationship of a type, as well as the const and volatile qualifications applied to types.

2.9.4 RTTI Layout

  1. The class definitions below are to be interpreted as implying a memory layout following the class layout rules for the host ABI. They specify data members only, except for the Standard-specified member functions of the std::type_info class given below, and do not imply anything about the member functions of these classes. Virtual member functions of these classes may only be used within the target systems' respective runtime libraries. The data members must be laid out exactly as specified.

  2. Every virtual table shall contain one entry describing the offset from a virtual pointer for that virtual table to the origin of the object containing that virtual pointer (or equivalently: to the virtual pointer for the primary virtual table). This entry is directly useful to implement dynamic_cast<void cv*>, but is also needed for the other truly dynamic casts. This entry is located two words ahead of the location pointed to by the virtual pointer (i.e., entry "-2"). This entry is present in all virtual tables, even for classes having virtual bases but no virtual functions.

  3. Every virtual table shall contain one entry that is a pointer to an object derived from std::type_info. This entry is located at the word preceding the location pointed to by the virtual pointer (i.e., entry "-1"). The entry is allocated in all virtual tables; for classes having virtual bases but no virtual functions, the entry is zero.

    We add one pointer to the std::type_info class in addition to the virtual table pointer implied by its virtual destructor:

    
          class type_info {
    	 ... // See section 2.9.3
    	private:
    	 const char *__type_name;
          };
    
    

    The possible derived types are:

  4. abi::__fundamental_type_info adds no data members to std::type_info;

  5. abi::__array_type_info and abi::__function_type_info do not add data members to std::type_info (these types are only produced by the typeid operator; they decay in other contexts).   abi::__enum_type_info does not add data members either.

  6. Three different types are used to represent user type information:

    1. abi::__class_type_info is used for class types having no bases, and is also a base type for the other two class type representations.
      
            class __class_type_info : public std::type_info {}
      
      

      This RTTI class may also be used for incomplete class types when referenced by a pointer RTTI, in which case it must be prevented from preempting the RTTI for the complete class type, for instance by emitting it as a static object (without external linkage).

      Two abi::__class_type_info objects can always be compared, for equality (i.e. of the types represented) or ordering, by comparison of their name NTBS addresses. In addition, complete class RTTI objects may also be compared for equality by comparison of their type_info addresses.

    2. For classes containing only a single, public, non-virtual base at offset zero (i.e. the derived class is dynamic iff the base is), class abi::__si_class_type_info is used. It adds to abi::__class_type_info a single member pointing to the type_info structure for the base type, declared "__class_type_info const *__base_type".
      
            class __si_class_type_info : public __class_type_info {
      	public:
      	  const __class_type_info *__base_type;
            };
      
      

    3. For classes with bases that do not satisfy the __si_class_type_info constraints, abi::__vmi_class_type_info is used. It is derived from abi::__class_type_info:
      
            class __vmi_class_type_info : public __class_type_info {
      	public:
      	  unsigned int __flags;
      	  unsigned int __base_count;
      	  __base_class_type_info __base_info[1];
      
      	  enum __flags_masks {
      	    __non_diamond_repeat_mask = 0x1,
      	    __diamond_shaped_mask = 0x2
      	  };
            };
      
      

      • __flags is a word with flags describing details about the class structure, which may be referenced by using the __flags_masks enumeration.
        • 0x01: class has non-diamond repeated inheritance
        • 0x02: class is diamond shaped
        These flags refer to both direct and indirect bases. The type of the __flags field is defined by each psABI, but must be at least 16 bits. For the 64-bit Itanium ABI, it will be unsigned int (32 bits).

      • __base_count is a word with the number of direct proper base class descriptions that follow. The type of the __base_count field is defined by each psABI. For the 64-bit Itanium ABI, it will be unsigned int (32 bits).

      • __base_info[] is an array of base class descriptions -- one for every direct proper base. Each description is of the type:
        
              struct abi::__base_class_type_info {
        	public:
                 const __class_type_info *__base_type;
        	 long __offset_flags;
        
        	 enum __offset_flags_masks {
        	   __virtual_mask = 0x1,
        	   __public_mask = 0x2,
        	   __offset_shift = 8
        	 };
        
              };
        

        The __base_type member points to the RTTI for the base type.

        All but the lower 8 bits of __offset_flags are a signed offset. For a non-virtual base, this is the offset in the object of the base subobject. For a virtual base, this is the offset in the virtual table of the virtual base offset for the virtual base referenced (negative).

        The low-order byte of __offset_flags contains flags, as given by the masks from the enumeration __offset_flags_masks:

        • 0x1: Base class is virtual
        • 0x2: Base class is public

      Note that the resulting structure is variable-length, with the actual size depending on the number of trailing base class descriptions.

  7. abi::__pbase_type_info is a base for both pointer types and pointer-to-member types. It adds two data members:
    
          class __pbase_type_info : public std::type_info {
    	public:
    	  unsigned int __flags;
    	  const std::type_info *__pointee;
    
    	  enum __masks {
    	    __const_mask = 0x1,
    	    __volatile_mask = 0x2,
    	    __restrict_mask = 0x4,
    	    __incomplete_mask = 0x8,
    	    __incomplete_class_mask = 0x10,
    	    __transaction_safe_mask = 0x20
    	    __noexcept_mask = 0x40
    	  };
          };
    
    

    • __flags is a flag word describing the cv-qualification and other attributes of the type pointed to (e.g., "int volatile*" should have the "volatile" bit set in that word). For pointer to function and pointer-to-member-function types, __flags is also used to indicate a "qualification" of sorts for the pointed-to (member) function type. That is the case for __transaction_safe_mask and __noexcept_mask where __pointee is a pointer to the "unqualified" version of the function type (e.g., one without the exception specification in the case of __noexcept_mask).

    • __pointee is a pointer to the std::type_info derivation for the unqualified type being pointed to.

    Note that the __flags bits should not be folded into the pointer to allow future definition of additional flags. It contains the following bits, and may be referenced using the flags defined in the __masks enum:

    • 0x1: __pointee type has const qualifier
    • 0x2: __pointee type has volatile qualifier
    • 0x4: __pointee type has restrict qualifier
    • 0x8: __pointee type is incomplete
    • 0x10: class containing __pointee is incomplete (in pointer to member)
    • 0x20: __pointee type is function type without the transaction-safe indication
    • 0x40: __pointee type is function type without the exception specification

    When the abi::__pbase_type_info is for a direct or indirect pointer to an incomplete class type, the incomplete target type flag is set. When it is for a direct or indirect pointer to a member of an incomplete class type, the incomplete class type flag is set. In addition, it and all of the intermediate abi::__pointer_type_info structs in the chain down to the abi::__class_type_info for the incomplete class type must be prevented from resolving to the corresponding type_info structs for the complete class type, possibly by making them local static objects. Finally, a dummy class RTTI is generated for the incomplete type that will not resolve to the final complete class RTTI (because the latter need not exist), possibly by making it a local static object.

    Two abi::__pbase_type_info objects can always be compared for equality (i.e. of the types represented) or ordering by comparison of their name NTBS addresses. In addition, unless either or both have either of the incomplete flags set, equality can be tested by comparing the type_info addresses.

  8. abi::__pointer_type_info is derived from abi::__pbase_type_info with no additional data members.

  9. The abi::__pointer_to_member_type_info type adds one field to abi::__pbase_type_info:
    
          class __pointer_to_member_type_info : public __pbase_type_info {
    	public:
    	  const abi::__class_type_info *__context;
          };
    
    
    • __context is a pointer to an abi::__class_type_info corresponding to the class type containing the member pointed to (e.g., the "A" in "int A::*")

<b>NOTE</b>: Note that this ABI requires elsewhere that a virtual table be emitted for a dynamic type in the object where the first non-inline virtual function member is defined, if any, or everywhere referenced if none. Therefore, an implementation should include at least one non-inline virtual function member and define it in the library, to avoid having user code inadvertently preempt the virtual table. Since the Standard requires a virtual destructor, and it will rarely be called, it is a good candidate for this role.

2.9.5 std::type_info::name()

The null-terminated byte string returned by this routine is the mangled name of the type.

2.9.6 The dynamic_cast Operator

Although dynamic_cast can work on pointers and references, from the point of view of representation we need only to worry about polymorphic class types. Also, some kinds of dynamic_cast operations are handled at compile time and do not need any RTTI. There are then three kinds of truly dynamic cast operations:

The most common kind of dynamic_cast is base-to-derived in a singly inherited hierarchy.

2.9.7 The dynamic_cast Algorithm

Dynamic casts to "void cv*" are inserted inline at compile time. So are dynamic casts of null pointers and dynamic casts that are really static.

This leaves the following test to be implemented in the run-time library for truly dynamic casts of the form "dynamic_cast<T>(v)": (see [expr.dynamic_cast] 5.2.7/8)

The first check corresponds to a "base-to-derived cast" and the second to a "cross cast". These tests are implemented by abi::__dynamic_cast:


   extern "C" 
   void* __dynamic_cast ( const void *sub,
			  const abi::__class_type_info *src,
			  const abi::__class_type_info *dst,
			  std::ptrdiff_t src2dst_offset);
   /* sub: source address to be adjusted; nonnull, and since the
    *      source object is polymorphic, *(void**)sub is a virtual
    pointer.
    * src: static type of the source object.
    * dst: destination type (the "T" in "dynamic_cast<T>(v)").
    * src2dst_offset: a static hint about the location of the
    *    source subobject with respect to the complete object;
    *    special negative values are:
    *       -1: no hint
    *       -2: src is not a public base of dst
    *       -3: src is a multiple public base type but never a
    *           virtual base type
    *    otherwise, the src type is a unique public nonvirtual
    *    base type of dst at offset src2dst_offset from the
    *    origin of dst.
    */

<b>NOTE</b>: Rationale:

2.9.8 The Exception Handler Matching Algorithm

Since the RTTI related exception handling routines are "personality specific", no interfaces need to be specified in this document (beyond the layout of the RTTI data).


Chapter 3: Code Emission and APIs


This chapter describes how to define and call functions. It also specifies the APIs of a variety of runtime utility routines required to be part of the support library of an ABI-conforming implementation for use by compiled code. In addition, reference is made to the separate description of exception handling in this ABI, which defines a large number of runtime utility routine APIs.

3.1 Functions

In general, the calling conventions and rules for defining C++ functions in this ABI follow those specified for functions of the corresponding type in the base C ABI. The corresponding type is mostly determined by translating C++ constructs to their obvious C analogues. This section specifies the behavior of of features without analogues in C, as well as some exceptions and extra rules required by C++-specific semantics.

3.1.1 Function Definitions

For the most part, non-static member functions, including constructors and destructors, are defined as if they were ordinary functions except for the addition of the implicit parameters to the prototype as described in the section on parameters.

The rules for member function pointers may require aligning the first instruction of ordinary non-static member functions (i.e. not constructors or destructors) to a higher value than the platform would normally require.

3.1.2 Parameters

3.1.2.1 this Parameters

Non-static member functions, including constructors and destructors, take an implicit this parameter of pointer type. It is passed as if it were the first parameter in the function prototype, except as modified for non-trivial return values.

3.1.2.2 VTT Parameters

Base-subobject constructors and destructors for classes with virtual bases take an implicit VTT parameter of pointer type. It is passed as if it were the second parameter in the function prototype, immediately following the this parameter, except as modified for non-trivial return values.

3.1.2.3 Non-Trivial Parameters

If a parameter type is a class type that is non-trivial for the purposes of calls, the caller must allocate space for a temporary and pass that temporary by reference. Specifically:

A C-style variadic argument of a type that is non-trivial for the purposes of calls is passed the same way: the address of the temporary is passed using the normal variadic mechanism, and va_arg in the callee retrieves the address and treats it as a reference to the temporary.

3.1.2.4 Parameters of Special Class Type

An argument of class std::decimal::decimal32, std::decimal::decimal64, or std::decimal::decimal128 as defined in TR 24733 is passed the same as the corresponding native decimal floating-point scalar type.

3.1.2.5 Reference Parameters

Reference parameters are handled by passing a pointer to the object bound to the reference.

3.1.2.6 Empty Parameters

Arguments of empty class types that are not non-trivial for the purposes of calls are passed no differently from ordinary classes.

On Itanium, the NaT bit must be set on all registers that are associated with the argument.

3.1.3 Return Values

3.1.3.1 Non-trivial Return Values

If the return type is a class type that is non-trivial for the purposes of calls, the caller passes an address as an implicit parameter. The callee then constructs the return value into this address. If the return type has a non-trivial destructor, the caller is responsible for destroying the temporary when control is returned to it normally. If an exception is thrown out of the callee after the return value is constructed but before control returns to the caller, e.g. by a throwing destructor, it is the callee's responsibility to destroy the return value before propagating the exception to the caller. Thus, in general, the caller is responsible for destroying the return value after, and only after, the callee returns control to the caller normally.

The address passed need not be of temporary memory; copy elision may cause it to point anywhere, including to global or heap-allocated memory.

C ABIs usually provide treatment for "indirect" return values, e.g. when returning a large aggregate that cannot fit in registers. In some cases, this treatment may not be suitable for non-trivial C++ return values, such as if the convention requires implicit copying or does not permit the return value to be constructed at an arbitrary address. If the treatment exists and is suitable, it is used for non-trivial return values. Otherwise, the pointer is passed as if it were the first parameter in the function prototype, preceding all other parameters, including the this and VTT parameters.

3.1.3.2 Return Values of Special Class Type

A return value of class std::decimal::decimal32, std::decimal::decimal64, or std::decimal::decimal128 as defined in TR 24733 is returned the same as the corresponding native decimal floating-point scalar type.

3.1.3.3 Reference Return Values

A return value of reference type is returned as a pointer to the object bound to the reference.

3.1.3.4 Empty Return Values

A return value of an empty class type that is not non-trivial for the purposes of calls will be returned as though it were the following C type:

struct { char c; };

On Itanium, the NaT bit must not be set for any register associated with this return value.

3.1.3.5 Constructor Return Values

Constructors return void results.

Some platforms are known to modify this rule to specify that constructors return a pointer to this. This may permit more efficient code generation in the caller.

3.1.3.6 Destructor Return Values

Destructors return void results.

Some platforms are known to modify this rule to specify that destructors return a pointer to this. This may permit more efficient code generation in the caller. This modified rule does not apply to deleting destructors. It also does not apply when making a virtual call to a complete-object destructor, so that this-adjustment thunks do not need to adjust the return value after the call.

3.2 Virtual Calls

3.2.1 General

This section sketches the calling convention for virtual functions, based on the above virtual table layout. See also the ABI examples document for motivating examples and potential implementations.

We explain, at a high level, what information must be present in the virtual table for a class A which declares a virtual function f in order that, given an pointer of type A*, the caller can call the virtual function f. This section does not specify exactly where that information is located (see above), nor does it specify how to convert a pointer to a class derived from A to an A*, if that is required.

When this section uses the term function pointer it is understood that this term may refer either to a traditional function pointer (i.e., a pointer to a GP/address pair) or a GP/address pair itself. Which of these alternatives is actually used is specified elsewhere in the ABI, but is independent of the description in this section.

Throughout this section, we assume that A is the class for which we are creating a virtual table, B is the most derived class in the hierarchy, and C is the class that contains C::f, the unique final overrider for A::f. This section specifies the contents of the f entry in the A-in-B virtual table. (If A is primary base in the hierarchy, then the A-in-B virtual table will be shared with the derived class virtual table -- but the contents of the A portion of that virtual table will still be as specified here.)

In all cases, the non-adjusting entry point for a virtual function expects the `this' pointer to point to an instance of the class in which the virtual function is defined. In other words, the non-adjusting entry point for C::f will expect that its `this' pointer points to a C object.

3.2.2 Virtual Table Components

For each virtual function declared in a class C, we add an entry to its virtual table if one is not already there (i.e. if it is not overriding a function in its primary base). In particular, a declaration which overrides a function inherited from a secondary base gets a new slot in the primary virtual table. We do this to avoid useless adjustments when calling a virtual function through a pointer to the most derived class.

The content of this entry for class A is a function pointer, as determined by one of the following cases. Recall that we are dealing with a hierarchy where B is most derived, A is a direct (or indirect) base of B defining f, and C contains the unique final overrider C::f of A::f.

  1. A = C

    (In this case, we are creating either the primary virtual table for A, or the A-in-B secondary virtual table.)

    The virtual table contains a function pointer pointing to the non-adjusting entry point for A::f.

  2. A != C

    In this case, we are creating the A-in-B secondary virtual table.

    The virtual table contains a pointer to an entry point that performs the adjustment from an A* to a C*, and then transfers control to the non-adjusting entry point for C::f.

There are some exceptions to this determination of function pointers:

When a class is used as a virtual base, we add a vcall offset slot to the beginning of its virtual table for each of the virtual functions it provides, whether in its primary or secondary virtual tables. Derived classes which override these functions may use the slots to determine the adjustment necessary.

3.2.3 Callee

For each direct or indirect base A of C that is not a morally virtual base of C, the compiler must emit, in the same object file as the code for C::f, an A-adjusting entry point for C::f. This entry point will expect that its this pointer points to an A*, and will convert it to a C* (which merely requires adding a constant offset) before transferring control to the non-adjusting entry point for C::f.

For each direct or indirect virtual base V of C such that V declares f, the compiler must emit, in the same object file as the code for C::f, a V-adjusting entry point for C::f. This entry point will expect that its this pointer points to the unique virtual V subobject of C. (Note that there may in general be multiple V subobjects of C, but that only one of them will be virtual.) This entry point must load the vcall offset corresponding to f located in the virtual table for V obtained via its this pointer, extract the vcall offset corresponding to f located in that virtual table, and add this offset to the this pointer. (Note that, as specified in the data layout document, when V is used as a virtual base, its virtual table contains vcall offsets for every virtual function declared in V or any of its bases.) Then, this entry point must transfer control to the non-adjusting entry point.

For each morally virtual base M of C such that M is not a virtual base (and therefore must be a subobject of a virtual base V), and such that M declares f, the compiler must emit, in the same object file as the code for C::f, an M-adjusting entry point for C::f. This entry point will expect that its this pointer points to an M*, and will convert it to a V* (a fixed offset), where V is the nearest virtual base to M along the inheritance path from C to M. Then, it will convert the V* to a C* by using the vcall offset stored in the V's virtual table.

3.2.4 Caller

When calling a virtual function f, through a pointer of static type B*, the caller

(Note that in general it will be optimal to select the class which contained the final overrider (i.e., C) as the class to which the B* should be converted. This class is always a satisfactory choice, since it is known to contain a definition of f. In addition, if the dynamic type of the object is B, then C::f will be the function ultimately selected by the call, which means that C's virtual table will contain a pointer to the non-adjusting entry point, meaning that no additional adjustments to the this pointer will be required.

However, there may be cases in which choosing a different base subobject could be superior. For example, if there is an alternate base D which also declares f, and a pointer to the D subobject is already available, then it may be better to use the D subobject rather than converting the B* to a C*, in order to avoid the cost of the conversion.)

3.2.5 Implementation

Note that the ABI only specifies the multiple entry points for a virtual function and its associated thunks; how those entry points are provided is unspecified. An existing compiler which uses thunks with a different means of adjusting the virtual table pointers can be made compliant with this ABI by only adding the vcall offsets -- the thunks need not use them. A more efficient implementation would be to emit all of the thunks immediately before the non-adjusting entry point to the function. Another might emit a new copy of the function for each entry point; this is a quality of implementation issue. See further discussion of implementation in the ABI examples document.

3.2.6 Pure Virtual Function API

An implementation shall provide a standard entry point that a compiler may reference in virtual tables to indicate a pure virtual function. Its interface is:

  extern "C" void __cxa_pure_virtual ();

This routine will only be called if the user calls a non-overridden pure virtual function, which has undefined behavior according to the C++ Standard. Therefore, this ABI does not specify its behavior, but it is expected that it will terminate the program, possibly with an error message.

3.2.7 Deleted Virtual Function API

An implementation shall provide a standard entry point that a compiler will reference in virtual tables to indicated a deleted virtual function. Its interface is:

  extern "C" void __cxa_deleted_virtual ();

This routine shall not return and while this ABI does not otherwise specify its behavior, it is expected that it will terminate the program, possibly with an error message.

3.3 Construction and Destruction APIs

This section describes APIs to be used for the construction and destruction of objects. This includes:

3.3.1 Subobject Construction and Destruction

The complete object constructors and destructors find the VTT, described in Section 2.6, Virtual Tables During Object Construction, via its mangled name. They pass the address of the subobject's sub-VTT entry in the VTT as a second parameter when calling the base object constructors and destructors. The base object constructors and destructors use the addresses passed to initialize the primary virtual pointer and virtual pointers that point to the classes which either have virtual bases or override virtual functions with a virtual step (have vcall offsets needing adjustment).

If a constructor calls constructors for base class subobjects that do not need construction virtual tables, e.g. because they have no virtual bases, the construction virtual table parameter is not passed to the base class subobject constructor, and the base class subobject constructors use their complete object virtual tables for initialization.

If a class has a non-virtual destructor, and a deleting destructor is emitted for that class, the deleting destructor must correctly handle the case that the this pointer is NULL. All other destructors, including deleting destructors for classes with a virtual destructor, may assume that the this pointer is not NULL.

Suppose we have a subobject class D that needs a construction virtual table, derived from a base B that needs a construction virtual table as part of D, and possibly from others that do not need construction virtual tables. Then the sub-VTT and constructor code for D would look like the following:

     // Sub-VTT for D (embedded in VTT for its derived class X):
     static vtable *__VTT__1D [1+n+m] =
	{ D primary vtable,
	  // The sub-VTT for B-in-D in X may have further structure:
	  B-in-D sub-VTT (n elements),
	  // The secondary virtual pointers for D's bases have elements
	  // corresponding to those in the B-in-D sub-VTT,
	  // and possibly others for virtual bases of D:
	  D secondary virtual pointer for B and bases (m elements) }; 

     D ( D *this, vtable **ctorvtbls )
     {
	// (The following will be unwound, not a real loop):
	for ( each base A of D ) {

	   // A "boring" base is one that does not need a ctorvtbl:
	   if ( ! boring(A) ) {
	     // Call subobject constructors with sub-VTT index
	     // if the base needs it -- only B in our example:
	      A ( (A*)this, ctorvtbls + sub-VTT-index(A) ); 

	   } else {
	     // Otherwise, just invoke the complete-object constructor:
	      A ( (A*)this );
	   }
	}

        // Initialize virtual pointer with primary ctorvtbls address
	// (first element):
        this->vptr = ctorvtbls+0;	// primary virtual pointer

	// (The following will be unwound, not a real loop):
	for ( each subobject A of D ) {
	
	   // Initialize virtual pointers of subobjects with ctorvtbls
	   // addresses for the bases 
	   if ( ! boring(A) ) {
	      ((A*)this)->vptr = ctorvtbls + 1+n + secondary-vptr-index(A);
		   // where n is the number of elements in the sub-VTTs
	    
	   } else {
	     // Otherwise, just use the complete-object vtable:
	      ((A *)this)->vptr = &(A-in-D vtable);
	   }
	}

        // Code for D constructor.
	...
      }

A test program for this can be found in the ABI Examples document.

3.3.2 Construction by Inherited Constructor

A constructor inherited from a base class can be used to initialize a derived class object, if it is explicitly inherited by a using declaration. Formally, such initialization is not performed by a derived class constructor, and instead the initialization expression itself directly initializes each base class (recursively, if the constructor is inherited from an indirect base class) and each non-static data member. To reduce code duplication, this ABI permits such initialization to be factored out into an inheriting constructor function.

An inheriting constructor describes the initialization that would be performed when a constructor inherited from a base class is selected to initialize a derived class, including the default-initialization of the other base classes and the non-static data members of the derived class. If the inheriting constructor is a base subobject constructor and the inherited constructor constructs a morally virtual base subobject, the inheriting constructor does not take any user-declared parameters; otherwise, it takes the same parameters as the inherited constructor. In all other respects, an inheriting constructors behaves the same as a constructor of the derived class. For example:

struct X { X(); };
struct A { A(int); };
struct B : A { using A::A; };
struct C : virtual B { using B::B; X x; };
struct D : C { using C::C; };
C c(0);
D d(0);

// The initializations of c and d behave as if they call C::Ccomplete A(int)(0) and D::Dcomplete A(int)(0):
D::Dcomplete A(int)(int n) : Bbase A(int)(n), Cbase A(int)() {}	// _ZN1DCI11AEi
C::Ccomplete A(int)(int n) : Bbase A(int)(n), x() {}           	// _ZN1CCI11AEi
C::Cbase A(int)() : /*no init for vbase B*/ x() {}             	// _ZN1CCI21AEi
B::Bbase A(int)(int n) : A(n) {}                               	// _ZN1BCI21AEi

Inheriting constructors are not permitted to make copies of their parameters when passing them to the inherited constructor. If it would not be possible to transparently forward all parameters from the inheriting constructor to the inherited constructor, an inheriting constructor cannot be used, and a different implementation technique (such as emitting the initialization inline) must be used instead.

3.3.3 One-time Construction API

As described in Section 2.8, certain objects with static storage duration have associated guard variables used to support the requirement that they be initialized exactly once, the first time the scope declaring them is entered. An implementation that does not anticipate supporting multi-threading may simply check the first byte (i.e., the byte with lowest address) of that guard variable, initializing if and only if its value is zero, and then setting it to a non-zero value.

However, an implementation intending to support automatically thread-safe, one-time initialization (as opposed to requiring explicit user control for thread safety) may make use of the following API functions:

extern "C" int __cxa_guard_acquire ( __int64_t *guard_object );

Returns 1 if the initialization is not yet complete; 0 otherwise. This function is called before initialization takes place. If this function returns 1, either __cxa_guard_release or __cxa_guard_abort must be called with the same argument. The first byte of the guard_object is not modified by this function.

A thread-safe implementation will probably guard access to the first byte of the guard_object with a mutex. If this function returns 1, the mutex will have been acquired by the calling thread.

extern "C" void __cxa_guard_release ( __int64_t *guard_object );

Sets the first byte of the guard object to a non-zero value. This function is called after initialization is complete.

A thread-safe implementation will release the mutex acquired by __cxa_guard_acquire after setting the first byte of the guard object.

extern "C" void __cxa_guard_abort ( __int64_t *guard_object );

This function is called if the initialization terminates by throwing an exception.

A thread-safe implementation will release the mutex acquired by __cxa_guard_acquire.

<b>NOTE</b>:

The following is pseudo-code showing how these functions can be used:

  if (obj_guard.first_byte == 0) {
    if ( __cxa_guard_acquire (&obj_guard) ) {
      try {
	... initialize the object ...;
      } catch (...) {
        __cxa_guard_abort (&obj_guard);
        throw;
      }
      ... queue object destructor with __cxa_atexit() ...;
      __cxa_guard_release (&obj_guard);
    }
  }

An implementation need not include the simple inline test of the initialization flag in the guard variable around the above sequence. If it does so, the cost of this scheme, when run single-threaded with minimal versions of the above functions, will be two extra function calls, each of them accessing the guard variable, the first time the scope is entered.

An implementation supporting thread-safety on multiprocessor systems must also guarantee that references to the initialized object do not occur before the load of the initialization flag. On Itanium, this can be done by using a ld1.acq operation to load the flag.

The intent of specifying an 8-byte structure for the guard variable, but only describing one byte of its contents, is to allow flexibility in the implementation of the API above. On systems with good small lock support, the second word might be used for a mutex lock. On others, it might identify (as a pointer or index) a more complex lock structure to use.

3.3.4 Array Construction and Destruction API

An ABI-compliant system shall provide several runtime routines for use in array construction and destruction. They may be used by compilers, but their use is not required. The required APIs are:

extern "C" void * __cxa_vec_new (
	    size_t element_count,
	    size_t element_size,
	    size_t padding_size,
	    void (*constructor) ( void *this ),
	    void (*destructor) ( void *this ) );

Equivalent to

  __cxa_vec_new2(element_count, element_size, padding_size, constructor,
                 destructor, &::operator new[], &::operator delete[])

extern "C" void * __cxa_vec_new2 (
	    size_t element_count,
	    size_t element_size,
	    size_t padding_size,
	    void (*constructor) ( void *this ),
	    void (*destructor) ( void *this ),
	    void* (*alloc) ( size_t size ),
	    void (*dealloc) ( void *obj ) );

Given the number and size of elements for an array and the non-negative size of prefix padding for a cookie, allocate space (using alloc) for the array preceded by the specified padding, initialize the cookie if the padding is non-zero, and call the given constructor on each element. Return the address of the array proper, after the padding.

If alloc throws an exception, rethrow the exception. If alloc returns NULL, return NULL. If the constructor throws an exception, call destructor for any already constructed elements, and rethrow the exception. If the destructor throws an exception, call std::terminate.

The constructor may be NULL, in which case it must not be called. If the padding_size is zero, the destructor may be NULL; in that case it must not be called.

Neither alloc nor dealloc may be NULL.

If the computed size of the allocated array object (including space for a cookie, if specified) would exceed the implementation-defined limit, std::bad_array_new_length is thrown.

extern "C" void * __cxa_vec_new3 (
	    size_t element_count,
	    size_t element_size,
	    size_t padding_size,
	    void (*constructor) ( void *this ),
	    void (*destructor) ( void *this ),
	    void* (*alloc) ( size_t size ),
	    void (*dealloc) ( void *obj, size_t size ) );
Same as __cxa_vec_new2 except that the deallocation function takes both the object address and its size.
extern "C" void __cxa_throw_bad_array_new_length (void);
Unconditionally throws std::bad_array_new_length. May be invoked by the compiler when the number of array elements expression of a new[] operation violates the requirements of the C++ standard.
extern "C" void __cxa_vec_ctor (
	    void *array_address,
	    size_t element_count,
	    size_t element_size,
	    void (*constructor) ( void *this ),
	    void (*destructor) ( void *this ) );
Given the (data) address of an array, not including any cookie padding, and the number and size of its elements, call the given constructor on each element. If the constructor throws an exception, call the given destructor for any already-constructed elements, and rethrow the exception. If the destructor throws an exception, call terminate(). The constructor and/or destructor pointers may be NULL. If either is NULL, no action is taken when it would have been called.
extern "C" void __cxa_vec_dtor (
	    void *array_address,
	    size_t element_count,
	    size_t element_size,
	    void (*destructor) ( void *this ) );
Given the (data) address of an array, the number of elements, and the size of its elements, call the given destructor on each element. If the destructor throws an exception, rethrow after destroying the remaining elements if possible. If the destructor throws a second exception, call terminate(). The destructor pointer may be NULL, in which case this routine does nothing.
extern "C" void __cxa_vec_cleanup (
	    void *array_address,
	    size_t element_count,
	    size_t element_size,
	    void (*destructor) ( void *this ) );
Given the (data) address of an array, the number of elements, and the size of its elements, call the given destructor on each element. If the destructor throws an exception, call terminate(). The destructor pointer may be NULL, in which case this routine does nothing.
extern "C" void __cxa_vec_delete (
	    void *array_address,
	    size_t element_size,
	    size_t padding_size,
	    void (*destructor) ( void *this ) );

If the array_address is NULL, return immediately. Otherwise, given the (data) address of an array, the non-negative size of prefix padding for the cookie, and the size of its elements, call the given destructor on each element, using the cookie to determine the number of elements, and then delete the space by calling ::operator delete[](void *). If the destructor throws an exception, rethrow after (a) destroying the remaining elements, and (b) deallocating the storage. If the destructor throws a second exception, call terminate(). If padding_size is 0, the destructor pointer must be NULL. If the destructor pointer is NULL, no destructor call is to be made.

<b>NOTE</b>: The intent of this function is to permit an implementation to call this function when confronted with an expression of the form delete[] p in the source code, provided that the default deallocation function can be used. Therefore, the semantics of this function are consistent with those required by the standard. The requirement that the deallocation function be called even if the destructor throws an exception derives from the resolution to DR 353 to the C++ standard, which was adopted in April, 2003.

extern "C" void __cxa_vec_delete2 (
	    void *array_address,
	    size_t element_size,
	    size_t padding_size,
	    void (*destructor) ( void *this ),
	    void (*dealloc) ( void *obj ) );
Same as __cxa_vec_delete, except that the given function is used for deallocation instead of the default delete function. If dealloc throws an exception, the result is undefined. The dealloc pointer may not be NULL.
extern "C" void __cxa_vec_delete3 (
	    void *array_address,
	    size_t element_size,
	    size_t padding_size,
	    void (*destructor) ( void *this ),
	    void (*dealloc) ( void *obj, size_t size ) );
Same as __cxa_vec_delete, except that the given function is used for deallocation instead of the default delete function. The deallocation function takes both the object address and its size. If dealloc throws an exception, the result is undefined. The dealloc pointer may not be NULL.
extern "C" void __cxa_vec_cctor (
	    void *dest_array,
	    void *src_array,
	    size_t element_count,
	    size_t element_size,
	    void (*constructor) (void *destination, void *source),
	    void (*destructor) (void *));
Given the (data) addresses of a destination and a source array, an element count and an element size, call the given copy constructor to copy each element from the source array to the destination array. The copy constructor's arguments are the destination address and source address, respectively. If an exception occurs, call the given destructor (if non-NULL) on each copied element and rethrow. If the destructor throws an exception, call terminate(). The constructor and or destructor pointers may be NULL. If either is NULL, no action is taken when it would have been called.

3.3.5 Controlling Object Construction Order

3.3.5.1 Motivation

The only requirement of the C++ Standard with respect to file scope object construction order is that file scope objects in a single object file are constructed in declaration order. However, building large programs sometimes requires careful attention to construction ordering for objects in different object files, and a number of vendors have provided extra-lingual facilities to control it. This ABI does not require an implementation to support this capability, but it specifies such a facility for those implementations that do.

This facility only controls construction order within a singled linked object (executable or DSO). Construction order between linked objects is determined by the initialization ordering specified in the base ABI.

3.3.5.2 Source Code API

A user may specify the construction priority with the pragma:

    #pragma priority ( <priority> )
The <priority> parameter specifies a 32-bit signed initialization priority, with lower numbers meaning earlier initialization. The range of priorities [MIN_INT .. MIN_INT+1023] is reserved to the implementation. The pragma applies to all file scope variables in the file where it appears, from the point of appearance to the next priority pragma or the end of the file. Objects defined before any priority pragmas have a default priority of zero, as do initialization actions specified by other means, e.g. DT_INIT_ARRAY entries. For consistency with the C++ Standard requirements on initialization order, behavior is undefined unless the priorities appearing in a single file, including any default zero priorities, are in non-decreasing numeric (non-increasing priority) order.

Initialization entries with the same priority from different files (or from other sources such as link command options) will be executed in an unspecified order.

3.3.5.3 Object File Representation

Initialization priority is represented in the object file by elements of a target-specific section type, SHT_IA_64_PRIORITY_INIT, with section ID 0x79000000 on Itanium, and section name .priority_init, and attributes allowing writing but not execution. The elements are structs:

	typedef struct {
	  ElfXX_Word	pi_pri;
	  ElfXX_Addr	pi_addr;
	} ElfXX_Priority_Init;
The field pi_addr is a function pointer, as defined by the base ABI (a pointer to a function descriptor on Itanium). The function takes a single unsigned int priority parameter, which performs some initialization at priority pi_pri. The priority value is obtained from the signed int in the source pragma by subtracting MIN_INT, so the default priority is -MIN_INT. The section header field sh_entsize is 8 for ELF-32, or 16 for ELF-64.

<b>NOTE</b>: An implementation may initialize as many (or as few) objects of the same priority as it chooses in a single such initialization function, as long as the sequence of such initialization entries for a given file preserves the source code order of objects to be initialized.

3.3.5.4 Runtime Library Support

Each implementation supporting priority initialization shall provide a runtime library function with prototype:

    void __cxa_priority_init ( ElfXX_Priority_Init *pi, int cnt );
It will be called with the address of a cnt-element (sub-)vector of the priority initialization entries, and must call each of them in order. It will be called with the GP of the initialization entries.

3.3.5.5 Linker Processing

The only required static linker processing is to concatenate the SHT_IA_64_PRIORITY_INIT sections in link order, which, given equal section IDs, section names, and section attributes as specified above, is the default behavior specified by the generic ABI for unknown section types.

<b>NOTE</b>: Given minimum static linker processing, an implementation supporting priority initialization would need to include bracketing files in the link command that (1) label the ends of the SHT_IA_64_PRIORITY_INIT section, and (2) provide initial and final DT_INIT_ARRAY entries. The initial DT_INIT_ARRAY entry would need to sort the SHT_IA_64_PRIORITY_INIT section and call __cxa_priority_init to run the constructors with negative priority (in the source). The final DT_INIT_ARRAY entry would need to call __cxa_priority_init to run the constructors with non-negative priority. Other DT_INIT_ARRAY entries would thus run at the proper point in the priority sequence.

A more ambitious linker implementation could sort the SHT_IA_64_PRIORITY_INIT section at link time and fabricate the code to call __cxa_priority_init at the beginning and end. At the extreme, it could even include other DT_INIT_ARRAY entries in the SHT_IA_64_PRIORITY_INIT sequence at the appropriate places and emit exactly one call to __cxa_priority_init, with no other entries in the DT_INIT_ARRAY section.

3.3.6 DSO Object Destruction API

3.3.6.1 Motivation

The C++ Standard requires that destructors be called for global objects when a program exits in the opposite order of construction. Most implementations have handled this by calling the C library atexit routine to register the destructors. This is problematic because the 1999 C Standard only requires that the implementation support 32 registered functions, although most implementations support many more. More important, it does not deal at all with the ability in most implementations to remove DSOs from a running program image by calling dlclose prior to program termination.

The API specified below is intended to provide standard-conforming treatment during normal program exit, which includes executing atexit-registered functions in the correct sequence relative to constructor-registered destructors, and reasonable treatment during early DSO unload (e.g. dlclose).

3.3.6.2 Runtime Data Structure

The runtime library shall maintain a list of termination functions with the following information about each:

The representation of this structure is implementation defined. All references are via the API described below.

3.3.6.3 Runtime API

  1. Object construction:

    After constructing a global (or local static) object, that will require destruction on exit, a termination function is registered as follows:

    extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
    This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the call f(p) when DSO d is unloaded, before all such termination calls registered before this one. It returns zero if registration is successful, nonzero on failure.

    The registration function is not called from within the constructor.

  2. User atexit calls:

    When the user registers exit functions with atexit, they should be registered with NULL parameters and DSO handles, i.e.

    __cxa_atexit ( f, NULL, NULL );
    It is expected that implementations supporting both C and C++ will integrate this capability into the libc atexit implementation so that C-only DSOs will nevertheless interact with C++ programs in a C++-standard-conforming manner. No user interface to __cxa_atexit is supported, so the user is not able to register an atexit function with a parameter or a home DSO.

  3. Termination:

    When linking any DSO containing a call to __cxa_atexit, the linker should define a hidden symbol __dso_handle, with a value which is an address in one of the object's segments. (It does not matter what address, as long as they are different in different DSOs.) It should also include a call to the following function in the FINI list (to be executed first):

    extern "C" void __cxa_finalize ( void *d );
    The parameter passed should be &__dso_handle.

    Note that the above can be accomplished either by explicitly providing the symbol and call in the linker, or by implicitly including a relocatable object in the link with the necessary definitions, using a .fini_array section for the FINI call. Also, note that these can be omitted for an object with no calls to __cxa_atexit, but they can be safely included in all objects.

    When __cxa_finalize(d) is called, it should walk the termination function list, calling each in turn if d matches __dso_handle for the termination function entry. If d == NULL, it should call all of them. Multiple calls to __cxa_finalize shall not result in calling termination function entries multiple times; the implementation may either remove entries or mark them finished.

    When the main program calls exit, it must call any remaining __cxa_atexit-registered functions, either by calling __cxa_finalize(NULL), or by walking the registration list itself.

    Note that the destructors must be called by __cxa_finalize() in the opposite of the order in which they were enqueued by __cxa_atexit.

Since __cxa_atexit and __cxa_finalize must both manipulate the same termination function list, they must be defined in the implementation's runtime library, rather than in the individual linked objects.

3.4 Demangler API

Synopsis:

namespace abi {
  extern "C" char* __cxa_demangle (const char* mangled_name,
				   char* buf,
				   size_t* n,
				   int* status);
}

Behavior: The return value is a pointer to a null-terminated array of characters, the demangled name. Ambiguities are possible between extern "C" object names and type manglings, e.g. "i" may be either an object named "i" or the built-in "int" type. Such ambiguous arguments are assumed to be type manglings. If the user has a set of external names to demangle, they should check that the names are in fact mangled (that is, begin with "_Z") before passing them to __cxa_demangle.

If there is an error in demangling, the return value is a null pointer. The user can examine *status to find out what kind of error occurred. Meaning of error indications:

Memory management:


Chapter 4: Exception Handling


See Exception Handling document, currently just the base psABI-level material, and the HP exception handling working paper, 8 December 1999.


Chapter 5: Linkage and Object Files


5.1 External Names (a.k.a. Mangling)

5.1.1 General

This section specifies the mangling, i.e. encoding, of external names (external in the sense of being visible outside the object file where they occur). The encoding is formalized as a derivation grammar along with the explanatory text, in a modified BNF with the following conventions:

See the separate table summarizing the encoding characters used as terminals. Also see additional mangling examples in the separate ABI examples document.

In the various explanatory examples, we use Ret? for an unknown function return type (i.e. that is not given by the mangling), or Type? for an unknown data type.

Mangled names containing $ or . are reserved for private implementation use. Names produced using such extensions are inherently non-portable and should be given internal linkage where possible.

5.1.2 General Structure

Entities with C linkage and global namespace variables are not mangled. Mangled names have the general structure:


    <mangled-name> ::= _Z <encoding>
                   ::= _Z <encoding> . <vendor-specific suffix>
    <encoding> ::= <function name> <bare-function-type>
	       ::= <data name>
	       ::= <special-name>
Thus, a name is mangled by prefixing "_Z" to an encoding of its name, and in the case of functions its type (to support overloading). At this top level, function types do not have the special delimiter characters required when nested (see below). Furthermore, in the case of instances (or explicit specializations) of function templates and member function templates (but not ordinary member functions of class templates), the <bare-function-type> encoding is that of the type expressed in the template (i.e., one likely involving template parameters). The type is omitted for variables and static data members.

A <mangled-name> containing a period represents a vendor-specific version or portion of the entity named by the <encoding> prior to the first period. There is no restriction on the characters that may be used in the suffix following the period.

ABI mangling is designed to ensure that entities receive the same mangling if and only if they are the same entity according to the C++ standard's one-definition rule (ODR) and the various rules for declaration matching (such as [over.dcl] and [temp.over]. Those rules are quite complex, and they dictate the results of mangling, and so it should not be surprising that the mangling rules are also complex. The ABI must be closely involved with the evolution of those language rules to ensure that they remain implementable with mangling. When the rules say that an ODR violation has undefined behavior, that is often because it is impractical to ensure that the entities involved will have different manglings. Similarly, when the rules forbid certain constructs from the signature of a declaration, that is often because that construct would create unreasonable problems for mangling.

Mangling must sometimes be able to distinguish entities that are not equivalent under the ODR and declaration-matching rules. This is true even if the entities would not be distinguishable by C++ code because, say, every name lookup which included both of them would be ambiguous. For example, different translation units might declare similar but not eqivalent function templates in the same namespace:

// a.cpp:
template <int> void foo() {}
template <> void foo<0>();

// b.cpp:
template <long> void foo() {}
template <> void foo<0>();

The C++ standard grants implementations broad flexibility to ignore certain kinds of differences. For example, the rules in [temp.over.link] for functionally-equivalent function templates could be used to shorten manglings in certain cases where instantiation-dependence provably has no effect. This ABI generally does not take advantage of that flexibility.

Dependent constructs in templates

It is sometimes necessary to mangle unresolved and uninstantiated language constructs such as types and expressions that appear within templates. This accounts for a lot of the complexity of entity mangling in this ABI.

In many places, the mangling grammar formally allows a single construct to be mangled in one of several different ways. Usually there is one production which allows a fully-resolved value or entity reference, and there is another production that allows an expression or unresolved entity reference. As an example, this can be clearly seen in the mangling for array types, which gives one mangling for a constant bound and another for an expression.

There are two reasons for this. First, manglings using the fully-resolved case are often significantly more compact. More importantly, though, the language often treat dependent and non-dependent constructs differently. For example, [temp.over.link] gives rules for when two expressions that involve template parameters are considered equivalent, and those rules are reflected in this ABI's expression mangling rules. Conversely, expressions that don't involve template parameters but are used in constant-evaluated contexts (such an array length) are considered to be equivalent if and only if they resolve to the same value. Mangling a non-dependent expression using its expression structure could incorrectly produce different manglings for different expressions that resolve to the same value, and it could incorrectly produce the same mangling for expressions that resolve to different values but happen to be spelled the same.

It is therefore important to use the right production given the dependence of the construct in question. The standard defines several different kinds of dependence, such as value dependence and type dependence. In general, the rule that should be used in mangling is instantiation dependence: if a construct in instantiation-dependent, it should use the general production, and otherwise it should use the narrow production. The grammar below will state clearly when certain productions are only for instantiation-dependent cases.

Anonymous entities

For the purposes of mangling, the name of an anonymous union is considered to be the name of the first named data member found by a pre-order, depth-first, declaration-order walk of the data members of the anonymous union. If there is no such data member (i.e., if all of the data members in the union are unnamed), then there is no way for a program to refer to the anonymous union, and there is therefore no need to mangle its name.

All of these examples:

union { int i; int j; };
union { union { int : 7 }; union { int i; }; };
union { union { int j; } i; };
are considered to have the name i for the purposes of mangling.

Names

In general, the mangling of an entity's name depends on where it is declared. Entities declared at global scope, or in namespace std, are mangled as unscoped names. Entities declared within a function, including members of local classes, are mangled with <local-name>. Entities declared in a namespace or class scope are mangled with <nested-name>. When the actual entity is not known statically, as can occur in a dependent function template signature, the name is mangled with <unresolved-name>.

The manglings of template specializations and non-template entities closely overlap, but they can generally be disambiguated by whether the name is followed by the I which starts a <template-args> production.


    <name> ::= <nested-name>
	   ::= <unscoped-name>
	   ::= <unscoped-template-name> <template-args>
	   ::= <local-name>	# See Scope Encoding below

    <unscoped-name> ::= <unqualified-name>
		    ::= St <unqualified-name>   # ::std::

    <unscoped-template-name> ::= <unscoped-name>
			     ::= <substitution>

A <nested-name> recursively breaks down the enclosing scope until the global scope is reached. A <prefix> refers to a scope; confusingly, a <template-prefix> actually refers to a template name (without template arguments).

Class and namespace members are always mangled with a <nested-name>, even if they are template specializations and there is an existing substitution for the template (and therefore the name could in principle be mangled as if it were a <unscoped-template-name>).

When a <nested-name> refers to a non-static class member function, the CV-qualifiers and ref-qualifiers of the function are prefixed to the compound name. This prefix is required even when the member function is a specialization of a substituted template and therefore those qualifiers could be inferred from the substitution target.


    <nested-name> ::= N [<CV-qualifiers>] [<ref-qualifier>] <prefix> <unqualified-name> E
		  ::= N [<CV-qualifiers>] [<ref-qualifier>] <template-prefix> <template-args> E

    <prefix> ::= <unqualified-name>                 # global class or namespace
             ::= <prefix> <unqualified-name>        # nested class or namespace
	     ::= <template-prefix> <template-args>  # class template specialization
             ::= <closure-prefix>                   # initializer of a variable or data member
             ::= <template-param>                   # template type parameter
             ::= <decltype>                         # decltype qualifier
	     ::= <substitution>

    <template-prefix> ::= <template unqualified-name>           # global template
                      ::= <prefix> <template unqualified-name>  # nested template
                      ::= <template-param>                      # template template parameter
                      ::= <substitution>

    <unqualified-name> ::= <operator-name> [<abi-tags>]
                       ::= <ctor-dtor-name>  
                       ::= <source-name>   
                       ::= <unnamed-type-name>   
                       ::= DC <source-name>+ E      # structured binding declaration

    <source-name> ::= <positive length number> <identifier>
    <identifier> ::= <unqualified source code identifier>

<identifier> is a pseudo-terminal representing the characters in the unqualified identifier for the entity in the source code. This ABI does not yet specify a mangling for identifiers containing characters outside of _A-Za-z0-9.

Note that <source-name> in the productions for <unqualified-name> may be either a function or data object name when derived from <name>, or a class or enum name when derived from <type>.

ABI tags

The GNU abi_tag attribute can be applied to a variable, function, inline namespace, class, or enumeration. The <unqualified-name> for a tagged variable, function, or type includes a representation of the tags on that entity, in alphabetical order:


    <abi-tags> ::= <abi-tag> [<abi-tags>]
    <abi-tag> ::= B <source-name>
For example:
  struct [[gnu::abi_tag ("foo","bar")]] A { }; // mangles as 1AB3barB3foo
If a name that would use a built-in <substitution> has ABI tags, the tags are appended to the substitution; the result is a substitutable component.
  namespace std
  {
    template <class T> struct char_traits { /* ... */ };
    template <class T> struct allocator { /* ... */ };
    template <class T, class R = char_traits<T>, class A = allocator<T>>
      struct [[gnu::abi_tag ("X")]] basic_string { /* ... */ };
    using string = basic_string<char>;
  }

  void f(std::string, std::string) { } // mangles as _Z1fSsB1XS_
If part of a declaration's type is not represented in the mangling, i.e. the type of a variable or a return type that is not represented in the mangling of a function, any ABI tags on that type (or components of a compound type) that are not also present in a mangled part of the type are applied to the name of the declaration. Note that there is no similar tag propagation from members or bases to a class type, as that would be impossible for incomplete types.

Note that for member functions of a class template that are not member templates, the type in question is that of the instantiation, so tags that appear only in the template do not affect mangling:

  struct [[gnu::abi_tag ("foo")]] A
  {
    template <class T> static T f();
    template <class T> static A g();
  };

  template <class T> struct B
  {
    static decltype(A::f<T>()) fa(decltype(A::f<T>()));
    static decltype(A::f<T>()) fv();
    static decltype(A::g<T>()) ga(decltype(A::g<T>()));
    static decltype(A::g<T>()) gv();
  };

  int main()
  {
    // decltype(A::f<T>()) resolves to int
    B<int>::fa(0);   // _ZN1BIiE2faEi
    B<int>::fv();    // _ZN1BIiE2fvEv
    // decltype(A::g<T>()) resolves to A, which has a tag
    B<int>::ga(A()); // _ZN1BIiE2gaE1AB3foo
    B<int>::gv();    // _ZN1BIiE2gvB3fooEv
  }

If no arguments are specified for the attribute on an inline namespace, the namespace has its own name as a tag. Tags on an inline namespace are not represented in the mangled name of the namespace, but they are subject to the above tag propagation. For example:

  inline namespace [[gnu::abi_tag]] Foo {
    struct A {};
    A f() { } // mangles as _ZN3Foo1fEv
  }
  template <class T> struct B { };
  typedef void (*fp)(B<A>);
  fp p;      // mangles as _Z1pB3Foo
  A g(A) { } // mangles as _Z1gN3Foo1AE
Numbers
    <number> ::= [n] <non-negative decimal integer>

<number> is a pseudo-terminal representing a decimal integer, with a leading 'n' for negative integers. It is used in <source-name> to provide the byte length of the following identifier. <number>s appearing in mangled names never have leading zeroes, except for the value zero, represented as '0'.

Sequence numbers
    <seq-id> ::= <0-9A-Z>+

A <seq-id> is a sequence number in base 36, using digits and upper case letters. Generally, wherever <seq-id> appears, the first element is encoded by the absence of a number, and the remainder of the sequence is encoded starting at 0. As with <number>, a <seq-id> has a leading zero only if that is the only digit.

For example, substitutions are mangled as S [<seq-id>] _. The first substitutable entity is encoded as S_, i.e. with no number. The second is encoded as S0_, the third as S1_, the twelfth as SA_, the thirty-eighth as S10_, etc.

5.1.3 Operator Encodings

Operators appear as function names, and in nontype template argument expressions. Unlike Cfront, unary and binary operators using the same symbol have different encodings. Most operators are encoded using exactly two letters, the first of which is lowercase.


  <operator-name> ::= nw	# new           
		  ::= na	# new[]
		  ::= dl	# delete        
		  ::= da	# delete[]      
		  ::= aw	# co_await      
		  ::= ps        # + (unary)
		  ::= ng	# - (unary)     
		  ::= ad	# & (unary)     
		  ::= de	# * (unary)     
		  ::= co	# ~             
		  ::= pl	# +             
		  ::= mi	# -             
		  ::= ml	# *             
		  ::= dv	# /             
		  ::= rm	# %             
		  ::= an	# &             
		  ::= or	# |             
		  ::= eo	# ^             
		  ::= aS	# =             
		  ::= pL	# +=            
		  ::= mI	# -=            
		  ::= mL	# *=            
		  ::= dV	# /=            
		  ::= rM	# %=            
		  ::= aN	# &=            
		  ::= oR	# |=            
		  ::= eO	# ^=            
		  ::= ls	# <<            
		  ::= rs	# >>            
		  ::= lS	# <<=           
		  ::= rS	# >>=           
		  ::= eq	# ==            
		  ::= ne	# !=            
		  ::= lt	# <             
		  ::= gt	# >             
		  ::= le	# <=            
		  ::= ge	# >=            
		  ::= ss	# <=>           
		  ::= nt	# !             
		  ::= aa	# &&            
		  ::= oo	# ||            
		  ::= pp	# ++ (postfix in <expression> context)
		  ::= mm	# -- (postfix in <expression> context)           
		  ::= cm	# ,             
		  ::= pm	# ->*           
		  ::= pt	# ->            
		  ::= cl	# ()            
		  ::= ix	# []            
		  ::= qu	# ?             
		  ::= cv <type>	# (cast)
                  ::= li <source-name>          # operator ""
		  ::= v <digit> <source-name>	# vendor extended operator

Vendors who define builtin extended operators (e.g. __imag) shall encode them as a 'v' prefix followed by the operand count as a single decimal digit, and the name in <length,ID> form.

<b>NOTE</b>: For a user-defined conversion operator the result type (i.e., the type to which the operator converts) is part of the mangled name of the function. If the conversion operator is a member template, the result type will appear before the template parameters. There may be forward references in the result type to the template parameters.

5.1.4 Other Special Functions and Entities

5.1.4.1 Virtual Tables and RTTI

Associated with a virtual table are several entities with mangled external names: the virtual table itself, the VTT for construction, the typeinfo structure, and the name it references. Each has a <special-name> encoding that is a simple two-character code, prefixed to the type encoding for the class to which it applies.


  <special-name> ::= TV <type>	# virtual table
		 ::= TT <type>	# VTT structure (construction vtable index)
		 ::= TI <type>	# typeinfo structure
		 ::= TS <type>	# typeinfo name (null-terminated byte string)

5.1.4.2 Virtual Override Thunks

Virtual function override thunks come in two forms. Those overriding from a non-virtual base, with fixed this adjustments, use a "Th" prefix and encode the required adjustment offset, probably negative, indicated by a 'n' prefix, and the encoding of the target function. Those overriding from a virtual base must encode two offsets after a "Tv" prefix. The first is the constant adjustment to the nearest virtual base (of the full object), of which the defining object is a non-virtual base. It is coded like the non-virtual case, with a 'n' prefix if negative. The second offset identifies the vcall offset in the nearest virtual base, which will be used to finish adjusting this to the full object. After these two offsets comes the encoding of the target function. The target function encodings of both thunks incorporate the function type; no additional type is encoded for the thunk itself.


  <special-name> ::= T <call-offset> <base encoding>
		      # base is the nominal target function of thunk
  <call-offset> ::= h <nv-offset> _
		::= v <v-offset> _
  <nv-offset> ::= <offset number>
		      # non-virtual base override
  <v-offset>  ::= <offset number> _ <virtual offset number>
		      # virtual base override, with vcall offset

Virtual function override thunks with covariant returns are twice as complex. Just as normal virtual function override thunks must adjust the this pointer before calling the base function, those with covariant returns must adjust the return pointer after they return from the base function. So the mangling must also encode a fixed offset to a non-virtual base, and possibly an offset to a vbase offset in the vtable to get to the virtual base containing the result subobject. We achieve this by encoding two <call-offset> components, either of which may be either virtual or non-virtual.


  <special-name> ::= Tc <call-offset> <call-offset> <base encoding>
		      # base is the nominal target function of thunk
		      # first call-offset is 'this' adjustment
		      # second call-offset is result adjustment

5.1.4.3 Constructors and Destructors

Constructors and destructors are simply special cases of <unqualified-name>, where the final <unqualified-name> of a nested name is replaced by one of the following:


  <ctor-dtor-name> ::= C1			# complete object constructor
		   ::= C2			# base object constructor
		   ::= C3			# complete object allocating constructor
		   ::= CI1 <base class type>	# complete object inheriting constructor
		   ::= CI2 <base class type>	# base object inheriting constructor
		   ::= D0			# deleting destructor
		   ::= D1			# complete object destructor
		   ::= D2			# base object destructor

The <base class type> in an inheriting constructor mangling identifies the base class in which the inherited constructor was originally declared.

Some of the symbols for constructor and destructor variants are optional.

5.1.4.4 Guard Variables

Initialization of certain objects with static storage duration requires a guard variable to prevent multiple initialization. The mangled name of a guard variable is the name of the guarded variable prefixed with GV.


  <special-name> ::= GV <object name>	# Guard variable for one-time initialization
			# No <type>

5.1.4.5 Lifetime-Extended Temporaries

The initializers of objects with static storage duration may introduce temporaries whose lifetime is extended to have static storage duration; this may also apply recursively to the initializers of those temporaries. If an initializer is visible to multiple translation units, those translation units must agree on the addresses of the temporaries. Therefore the temporaries must be given a consistent name and vague linkage. The mangled name of a temporary is the name of the non-temporary object in whose initializer they appear, prefixed with GR and suffixed with a sequence number mangled using the usual rules for a seq-id. Temporaries are numbered with a pre-order, depth-first, left-to-right walk of the complete initializer.


  <special-name> ::= GR <object name> _             # First temporary
  <special-name> ::= GR <object name> <seq-id> _    # Subsequent temporaries

For example, consider the following code:

struct A { const int (&x)[3]; };
struct B { const A (&x)[2]; };
template <typename T> B &&b = { { { { 1, 2, 3 } }, { { 4, 5, 6 } } } };
B &temp = b<void>;

5.1.4.6 Transaction-Safe Function Entry Points
A function declared transaction-safe or [[optimize_for_synchronized]] has two entry points: the normal function mangling, used for calls from a non-transaction context, and another entry point used for calls during a transaction. The mangled name of the transaction entry point is the normal mangling prefixed with GTt.

  <special-name> ::= GTt <encoding>

5.1.5 Type encodings

Types are encoded according to their compound structure: the tree of type constructors, such as const and *, that uniquely determine the type. The mangling of function template signatures necessitates the ability to encode the compound structure of dependent types.

Simple forms of type structure, such as reference and pointer types, are encoded with a single-character prefix. More complex forms of type structure, such as qualifiers and function types, require individual discussion below.


  <type> ::= <builtin-type>
         ::= <qualified-type>
         ::= <function-type>
         ::= <class-enum-type>
         ::= <array-type>
         ::= <pointer-to-member-type>
         ::= <template-param>
         ::= <template-template-param> <template-args>
         ::= <decltype>
         ::= P <type>        # pointer
         ::= R <type>        # l-value reference
         ::= O <type>        # r-value reference (C++11)
         ::= C <type>        # complex pair (C99)
         ::= G <type>        # imaginary (C99)
         ::= <substitution>  # See Compression below
5.1.5.1 Qualified types

  <qualified-type>     ::= <qualifiers> <type>

  <qualifiers>         ::= <extended-qualifier>* <CV-qualifiers>
  <extended-qualifier> ::= U <source-name> [<template-args>] # vendor extended type qualifier
  <CV-qualifiers>      ::= [r] [V] [K] 	  # restrict (C99), volatile, const

  <ref-qualifier>      ::= R              # & ref-qualifier
  <ref-qualifier>      ::= O              # && ref-qualifier

Vendors who define extended type qualifiers (e.g. _near and _far for pointers) shall encode them as a 'U' prefix, followed by the name in <length,ID> form, followed optionally by any arguments to the qualifier. It is recommended that the encoded name be the preferred name used in source code; known exceptions are listed below.

In cases where multiple order-insensitive qualifiers are present, they should be ordered (beginning closest to the base type) 'K', 'V', 'r', and 'U' (farthest from the base type), with the 'U' qualifiers in alphabetical order by the vendor name (with alphabetically earlier names closer to the base type). For example, int* volatile const restrict _far has mangled type name U4_farrVKPi. Vendors must therefore specify which of their extended qualifiers are considered order-insensitive. This need not necessarily be resolved on the basis of whether their language translators impose an order in source code. They are encouraged to resolve questionable cases as being order-insensitive to maximize consistency in mangling.

For purposes of substitution, given a qualified type, the base type is substitutible and the type with all the K, V, and r qualifiers plus any vendor extended types in the same order-insensitive set is substitutible; however, types with only a subset of those qualifiers are not. That is, given a type const volatile foo, the fully qualified type or foo may be substituted, but not volatile foo nor const foo.

<b>NOTE</b>: The restrict qualifier is part of the C99 standard, but is strictly an extension to C++ at this time. There is no standard specification of whether the restrict attribute is part of the type for overloading purposes. An implementation should include its encoding in the mangled name if and only if it also treats it as a distinguishing attribute for overloading purposes. This ABI does not specify that choice.

Known exceptions to the extended qualifier rules
5.1.5.2 Builtin types

Builtin types are represented by single-letter codes:


  <builtin-type> ::= v	# void
		 ::= w	# wchar_t
		 ::= b	# bool
		 ::= c	# char
		 ::= a	# signed char
		 ::= h	# unsigned char
		 ::= s	# short
		 ::= t	# unsigned short
		 ::= i	# int
		 ::= j	# unsigned int
		 ::= l	# long
		 ::= m	# unsigned long
		 ::= x	# long long, __int64
		 ::= y	# unsigned long long, __int64
		 ::= n	# __int128
		 ::= o	# unsigned __int128
		 ::= f	# float
		 ::= d	# double
		 ::= e	# long double, __float80
		 ::= g	# __float128
		 ::= z	# ellipsis
                 ::= Dd # IEEE 754r decimal floating point (64 bits)
                 ::= De # IEEE 754r decimal floating point (128 bits)
                 ::= Df # IEEE 754r decimal floating point (32 bits)
                 ::= Dh # IEEE 754r half-precision floating point (16 bits)
                 ::= DF <number> _ # ISO/IEC TS 18661 binary floating point type _FloatN (N bits), C++23 std::floatN_t
                 ::= DF <number> x # IEEE extended precision formats, C23 _FloatNx (N bits)
                 ::= DF16b # C++23 std::bfloat16_t
                 ::= DB <number> _        # C23 signed _BitInt(N)
                 ::= DB <instantiation-dependent expression> _ # C23 signed _BitInt(N)
                 ::= DU <number> _        # C23 unsigned _BitInt(N)
                 ::= DU <instantiation-dependent expression> _ # C23 unsigned _BitInt(N)
                 ::= Di # char32_t
                 ::= Ds # char16_t
                 ::= Du # char8_t
                 ::= Da # auto
                 ::= Dc # decltype(auto)
                 ::= Dn # std::nullptr_t (i.e., decltype(nullptr))
                 ::= [DS] DA  # N1169 fixed-point [_Sat] T _Accum
                 ::= [DS] DR  # N1169 fixed-point [_Sat] T _Fract
		 ::= u <source-name> [<template-args>] # vendor extended type

  <fixed-point-size>
                 ::= s # short
                 ::= t # unsigned short
                 ::= i # plain
                 ::= j # unsigned
                 ::= l # long
                 ::= m # unsigned long

Vendors who define builtin extended types shall encode them as a 'u' prefix followed by the name in <length,I> form, followed by any arguments to the extended type.

5.1.5.3 Function types

Function types are composed from their parameter types and possibly the result type. Except at the outer level type of an <encoding>, or in the <encoding> of an otherwise delimited external name in a <template-param> or <local-name> function encoding, these types are delimited by an "F..E" pair. For purposes of substitution (see Compression below), delimited and undelimited function types are considered the same.

Whether the mangling of a function type includes the return type depends on the context and the nature of the function. The rules for deciding whether the return type is included are:

  1. Template functions (names or types) have return types encoded, with the exceptions listed below.
  2. Function types not appearing as part of a function name mangling, e.g. parameters, pointer types, etc., have return type encoded, with the exceptions listed below.
  3. Non-template function names do not have return types encoded.
The exceptions mentioned in (1) and (2) above, for which the return type is never included, are

Empty parameter lists, whether declared as () or conventionally as (void), are encoded with a void parameter specifier (v). Therefore function types always encode at least one parameter type, and function manglings can always be distinguished from data manglings by the presence of the type. Member functions do not encode the types of implicit parameters, either this or the VTT parameter.

The mangling of CV-qualifiers and ref-qualifiers on a function type differs according to context. When mangling the name of a non-static member function, the CV-qualifiers and ref-qualifiers of that function are encoded at the beginning of the <nested-name> as described above. Otherwise, they are encoded as part of the function type as described below.

When an exception-specification (i.e., noexcept, noexcept(expression), or throw(type(s))) is part of the function type, it is mangled according to <exception-spec> as described below. A non-instantiation-dependent, potentially-throwing exception specification is not mangled.

A transaction-safe function type is encoded with a "Dx" before the "F". This affects only type mangling; a transaction-safe function has the same mangling as a non-transaction-safe function.

A "Y" prefix for the bare function type encodes extern "C" in implementations which distinguish between function types with "C" and "C++" language linkage. This affects only type mangling, since extern "C" function objects have unmangled names.


  <function-type> ::= [<CV-qualifiers>] [<exception-spec>] [Dx] F [Y] <bare-function-type> [<ref-qualifier>] E
  <bare-function-type> ::= <signature type>+
	# types are possible return type, then parameter types
  <exception-spec> ::= Do                # non-throwing exception-specification (e.g., noexcept, throw())
                   ::= DO <expression> E # computed (instantiation-dependent) noexcept
                   ::= Dw <type>+ E      # dynamic exception specification with instantiation-dependent types

For the purposes of substitution, the CV-qualifiers and ref-qualifier of a function type are an indivisible part of the type; that is, when mangling void () const, void () is not a substitution candidate.

When a function parameter is a C++11 function parameter pack, its type is mangled with Dp <type>, i.e., its type is a pack expansion:

 <type>  ::= Dp <type>          # pack expansion (C++11)

5.1.5.4 C++11 decltype

The C++11 decltype type is encoded with either Dt or DT, depending on how the decltype type was parsed. (See farther below for the encoding of expressions.)

 <decltype>  ::= Dt <expression> E  # decltype of an id-expression or class member access (C++11)
             ::= DT <expression> E  # decltype of an expression (C++11)
If the operand expression of decltype is not instantiation-dependent then the resulting type is encoded directly. For example:
          int x;
          template<class T> auto f(T p)->decltype(x);
            // The return type in the mangling of the template signature
            // is encoded as "i".
          template<class T> auto f(T p)->decltype(p);
            // The return type in the mangling of the template signature
            // is encoded as "Dtfp_E".
          void g(int);
          template<class T> auto f(T p)->decltype(g(p));
            // The return type in the mangling of the template signature
            // is encoded as "DTcl1gfp_E".
5.1.5.5 Class, union, and enum types

A class, union, or enum type is simply a name. It may be a simple <unqualified-name>, with or without a template argument list, or a more complex <nested-name>. Thus, it is encoded like a function name, except that no CV-qualifiers are present in a nested name specification.


  <class-enum-type> ::= <name>     # non-dependent type name, dependent type name, or dependent typename-specifier
                    ::= Ts <name>  # dependent elaborated type specifier using 'struct' or 'class'
                    ::= Tu <name>  # dependent elaborated type specifier using 'union'
                    ::= Te <name>  # dependent elaborated type specifier using 'enum'

An exception, however, is that class std::decimal::decimal32, std::decimal::decimal64, or std::decimal::decimal128 as defined in TR 24733 uses the same encoding as the corresponding native decimal-floating point scalar type.

Unnamed class, union, and enum types that aren't closure types, that haven't acquired a "name for linkage purposes" (through a typedef), and that aren't anonymous union types, follow the same rule when they are defined in class scopes, with the underlying <unqualified-name> an <unnamed-type-name> of the form

  <unnamed-type-name> ::= Ut [ <nonnegative number> ] _ 
The number is omitted for the first unnamed type in the class; it is n-2 for the nth unnamed type (in lexical order) otherwise.

(The mangling of such unnamed types defined in namespace scope is generally unspecified because they do not have to match across translation units. An implementation must only ensure that naming collisions are avoided. The mangling of such unnamed types in local scopes is described in Scope Encoding. The encoding of closure types is described in a Closure Types (Lambdas).)

For example:

	struct S { static struct {} x; };
	typedef decltype(S::x) TX;  // Type mangled as N1SUt_E
	TX S::x;                    // _ZN1S1xE
	void f(TX) {}               // _Z1fN1SUt_E
5.1.5.6 Array types

Array types encode their array bound and element type. Note that "array" parameters to functions are encoded as pointer types. The array bound (but not the _ separator) is omitted for incomplete array types (e.g. int[]) and C99 variable-length array types.


  <array-type> ::= A [<array bound number>] _ <element type>
	       ::= A <instantiation-dependent array bound expression> _ <element type>

The second rule is used when the array bound is an instantiation-dependent expression. For example:

    template<int I> void foo (int (&)[I + 1]) { }

    // Mangled as _Z3fooILi2EEvRAplT_Li1E_i
    template void foo<2> (int (&)[3]);
5.1.5.7 Pointer-to-member types

Pointer-to-member types encode the class and member types:

  <pointer-to-member-type> ::= M <class type> <member type>

For example,

    void f (void (A::*)() const &) {}
produces the mangled name "_Z1fM1AKFvvRE".
5.1.5.8 Template parameters

A reference to a template parameter is mangled using the index of the parameter, with a special mangling for the first parameter. The sequence of parameters is therefore T_, T0_, T1_, and so on.


  <template-param> ::= T_ # first template parameter
                   ::= T <parameter-2 non-negative number> _
  <template-template-param> ::= <template-param>
                            ::= <substitution>

For example:


    template<class T> void f(T) {}

    // Mangled as "_Z1fIiEvT_"
    template void f(int);

Note that a template parameter reference is a substitution candidate. As a substitution, it is treated as distinct from the actual template argument, including in recursive positions. For example, in the mangling of the following function template specialization, the first incidence of T* is not substituted despite being known (in this specialization) to be the same type as int*, and the second incidence is substituted with the substitution derived from the first incidence, not that from the incidence of int*.


    template<class T> void f(int*, T*, T*) {}

    // Mangled as "_Z1fIiEvPiPT_S2_"
    template void f(int*, int*, int*);

Typically, only references to function template parameters occurring within the dependent signature of the template are mangled this way. In other contexts, template instantiation replaces references to template parameters with the actual template arguments, and mangling should mangle such references exactly as if they were that template argument. For example:


    template<class T> class A {
      template<class U> void f(T, U) {}
    };

    // Mangled as "_ZN1AIiE1fIfEEviT_"
    template void A<int>::f(int, float);

5.1.5.9 Function parameter references

Function parameters referenced in other parameter types or in late-specified return types are handled similarly to template parameters, but involve a few more subtleties.

Let L be the number of function prototype scopes from the innermost one (in which the parameter reference occurs) up to (and including) the one containing the declaration of the referenced parameter. If the parameter declaration clause of the innermost function prototype scope has been completely seen, it is not counted (in that case -- which is perhaps the most common -- L can be zero). For example:

          template<class T> void f(T p, decltype(p));                         // L = 1
          template<class T> void g(T p, decltype(p) (*)());          // L = 1
          template<class T> void h(T p, auto (*)()->decltype(p));    // L = 1
          template<class T> void i(T p, auto (*)(T q)->decltype(q)); // L = 0
          template<class T> void j(T p, auto (*)(decltype(p))->T);   // L = 2
          template<class T> void k(T p, int (*(*)(T p))[sizeof(p)]); // L = 1


  <function-param> ::= fp <top-level CV-qualifiers> _                                     # L == 0, first parameter
		   ::= fp <top-level CV-qualifiers> <parameter-2 non-negative number> _   # L == 0, second and later parameters
		   ::= fL <L-1 non-negative number> p <top-level CV-qualifiers> _         # L > 0, first parameter
		   ::= fL <L-1 non-negative number> p <top-level CV-qualifiers>
                                                    <parameter-2 non-negative number> _   # L > 0, second and later parameters
		   ::= fpT                                                                # this
Note that top-level cv-qualifiers specified on a parameter type do not affect the function type directly (i.e., int(*)(T) and int(*)(T const) are the same type), but in expression contexts (such as decltype arguments) they do matter and must therefore be encoded in <function-param>, unless the parameter is used as an rvalue of a known non-class type (in the latter case the qualifier cannot affect the semantics of the expression). For example:
          template<typename T> void f(T const p, decltype(p)*);
            // The specialization f<int> has type void(int, int const*)
            // and is encoded as _Z1fIiEvT_PDtfL0pK_E
5.1.5.10 Template Arguments

Template argument lists appear after the unqualified template name, and are bracketed by I/E. This is used in names for specializations in particular, but also in types and scope identification. Template argument packs are bracketed by J/E to distinguish them from other arguments.


  <template-args> ::= I <template-arg>+ E

  <template-arg> ::= <type>                                             # type or template
                 ::= X <expression> E                                   # expression
                 ::= <expr-primary>                                     # simple expressions
                 ::= J <template-arg>* E                                # argument pack

Type arguments appear using their regular encoding. For example, the template class "A<char, float>" is encoded as "1AIcfE". A slightly more involved example is a dependent function parameter type "A<T2>::X" (T2 is the second template parameter) which is encoded as "N1AIT0_E1XE", where the "N...E" construct is used to describe a qualified name.

5.1.6 Expressions

Expressions must be mangled in several contexts.

When mangling the name of a specialized template, non-type template arguments are mangled as expressions. These expressions are typically very simple, and they do not necessarily reflect any argument expression that was used in source. See the section on mangling template arguments for more detail.

More generally, when mangling the signature of a function template, any instantiation-dependent expressions (e.g. in an array bound, decltype, or template argument) must be mangled in order to properly distinguish templates that are different under the ODR. See the section on dependent mangling. As a result, nearly the entire expression grammar of C++ is subject to mangling, with only a few exceptions (like lambdas) that are explicitly disallowed in function signatures.

In general, expression manglings reflect a prefix traversal of the syntactic expression tree, with parentheses omitted. (Parentheses may be ignored because they are implicit in the prefix representation and typically do not affect semantics. However, when parentheses are used to suppress argument-dependent lookup, the call expression may need to be mangled differently.) Unless explicitly stated otherwise, the expression is mangled without constant folding or other simplification.

Each expression mangling begins with a code (typically two letters) indicating the kind of expression, which dictates the form of the rest of the mangling. For overloadable operators, this code is the same as the <operator-name>.

For example, if J is the third template parameter, "B<(J+1)/2>" becomes "1BI Xdv pl T1_ Li1E Li2E E E" (the blanks are present only to visualize the decomposition).

If the operand of a sizeof or alignof operator is not instantiation-dependent, it is encoded as an integer literal reflecting the result of the operator. If the result of the operator is implicitly converted to a known integer type, that type is used for the literal; otherwise, the type of std::size_t or std::ptrdiff_t is used. For example:

          template<class T, int N> struct S1 {};
          template<class T, T N> struct S2 {};
          template<class T> void f(S1<T, sizeof(long double)>);
            // The sizeof(...) is not instantiation-dependent, and converted to int:
            // the result is encoded as "Li16E" for 16-byte long double types.
          template<class T> void f(S2<T, sizeof(long double)>);
            // The sizeof(...) is not instantiation-dependent, and converted to an
            // unknown type: the result is encoded as "Lm16E" for 16-byte long double
            // types and std::size_t a synonym for "unsigned long".
          template<class T> void f(S2<T, sizeof(T*)>);
            // The sizeof(...) is instantiation-dependent (even though its value may
            // be known if all pointers have the same size): It is encoded as "stPT_".


  <expression> ::= <unary operator-name> <expression>
               ::= <binary operator-name> <expression> <expression>
               ::= <ternary operator-name> <expression> <expression> <expression>
               ::= pp_ <expression>                                     # prefix ++
               ::= mm_ <expression>                                     # prefix --
               ::= cl <expression>+ E                                   # expression (expr-list), call
               ::= cp <base-unresolved-name> <expression>* E            # (name) (expr-list), call that would use argument-dependent lookup but for the parentheses
               ::= cv <type> <expression>                               # type (expression), conversion with one argument
               ::= cv <type> _ <expression>* E                          # type (expr-list), conversion with other than one argument
               ::= tl <type> <braced-expression>* E                     # type {expr-list}, conversion with braced-init-list argument
               ::= il <braced-expression>* E                            # {expr-list}, braced-init-list in any other context
               ::= [gs] nw <expression>* _ <type> E                     # new (expr-list) type
               ::= [gs] nw <expression>* _ <type> <initializer>         # new (expr-list) type (init)
               ::= [gs] na <expression>* _ <type> E                     # new[] (expr-list) type
               ::= [gs] na <expression>* _ <type> <initializer>         # new[] (expr-list) type (init)
               ::= [gs] dl <expression>                                 # delete expression
               ::= [gs] da <expression>                                 # delete[] expression
               ::= dc <type> <expression>                               # dynamic_cast<type> (expression)
               ::= sc <type> <expression>                               # static_cast<type> (expression)
               ::= cc <type> <expression>                               # const_cast<type> (expression)
               ::= rc <type> <expression>                               # reinterpret_cast<type> (expression)
               ::= ti <type>                                            # typeid (type)
               ::= te <expression>                                      # typeid (expression)
               ::= st <type>                                            # sizeof (type)
               ::= sz <expression>                                      # sizeof (expression)
               ::= at <type>                                            # alignof (type)
               ::= az <expression>                                      # alignof (expression)
               ::= nx <expression>                                      # noexcept (expression)
               ::= <template-param>
               ::= <function-param>
               ::= dt <expression> <unresolved-name>                    # expr.name
               ::= pt <expression> <unresolved-name>                    # expr->name
               ::= ds <expression> <expression>                         # expr.*expr
               ::= sZ <template-param>                                  # sizeof...(T), size of a template parameter pack
               ::= sZ <function-param>                                  # sizeof...(parameter), size of a function parameter pack
               ::= sP <template-arg>* E                                 # sizeof...(T), size of a captured template parameter pack from an alias template
               ::= sp <expression>                                      # expression..., pack expansion
               ::= fl <binary operator-name> <expression>               # (... operator expression), unary left fold
               ::= fr <binary operator-name> <expression>               # (expression operator ...), unary right fold
               ::= fL <binary operator-name> <expression> <expression>  # (expression operator ... operator expression), binary left fold
               ::= fR <binary operator-name> <expression> <expression>  # (expression operator ... operator expression), binary right fold
               ::= tw <expression>                                      # throw expression
               ::= tr                                                   # throw with no operand (rethrow)
               ::= u <source-name> <template-arg>* E                    # vendor extended expression
               ::= <unresolved-name>                                    # f(p), N::f(p), ::f(p),
                                                                        # freestanding dependent name (e.g., T::x),
                                                                        # objectless nonstatic member reference
               ::= <expr-primary>

  <unresolved-name> ::= [gs] <base-unresolved-name>                     # x or (with "gs") ::x
                    ::= sr <unresolved-type> <base-unresolved-name>     # T::x / decltype(p)::x
                    ::= srN <unresolved-type> <unresolved-qualifier-level>+ E <base-unresolved-name>
                                                                        # T::N::x /decltype(p)::N::x
                    ::= [gs] sr <unresolved-qualifier-level>+ E <base-unresolved-name>  
                                                                        # A::x, N::y, A<T>::z; "gs" means leading "::"

  <unresolved-type> ::= <template-param> [ <template-args> ]            # T:: or T<X,Y>::
                    ::= <decltype>                                      # decltype(p)::
                    ::= <substitution>

  <unresolved-qualifier-level> ::= <simple-id>

  <simple-id> ::= <source-name> [ <template-args> ]

  <base-unresolved-name> ::= <simple-id>                                # unresolved name
                         ::= on <operator-name>                         # unresolved operator-function-id
                         ::= on <operator-name> <template-args>         # unresolved operator template-id
                         ::= dn <destructor-name>                       # destructor or pseudo-destructor;
                                                                        # e.g. ~X or ~X<N-1>

  <destructor-name> ::= <unresolved-type>                               # e.g., ~T or ~decltype(f())
                    ::= <simple-id>                                     # e.g., ~A<2*N>

  <expr-primary> ::= L <type> <value number> E                          # integer literal
                 ::= L <type> <value float> E                           # floating literal
                 ::= L <string type> E                                  # string literal
                 ::= L <nullptr type> E                                 # nullptr literal (i.e., "LDnE")
                 ::= L <pointer type> 0 E                               # null pointer template argument
		 ::= L <type> <real-part float> _ <imag-part float> E   # complex floating point literal (C 2000)
                 ::= L _Z <encoding> E                                  # external name

  <braced-expression> ::= <expression>
                      ::= di <field source-name> <braced-expression>    # .name = expr
                      ::= dx <index expression> <braced-expression>     # [expr] = expr
                      ::= dX <range begin expression> <range end expression> <braced-expression>
                                                                        # [expr ... expr] = expr

  <initializer> ::= pi <expression>* E                                  # parenthesized initialization

A production for <expression> that directly specifies an operation code (e.g., for the -> operator) takes precedence over one that is expressed in terms of (unary/binary/ternary) <operator-name>.

The optional "gs" prefix on some of the productions indicates that the corresponding source construct (name, new-expression, or delete-expression) includes a global-scope qualifier (e.g., ::x).

tl is used for direct-list-initializations, where the type name is directly followed by a braced-init-list; e.g., MyArray{1,2,3} should be mangled tl7MyArrayLi1ELi2ELi3EE. If the braced-init-list is parenthesized, this is not a direct-list-initialization, and it should be mangled with cv and a nested il; for example, MyArray({1,2,3}) should be mangled cv7MyArrayilLi1ELi2ELi3EE.

If an implementation supports the full C99 designated initializer syntax (as an extension), a designator list comprising multiple designators results in multiple nested <braced-expression>s. For example, X{.a.b[3] = 1} should be mangled tl1Xdi1adi1bdxLi3ELi1EE.

In C++, a call expression where the callee operand is an unqualified name uses argument-dependent lookup unless unqualified lookup finds certain kinds of declarations; see C++11 [basic.lookup.argdep]p3. Because this rule does not apply when the name is parenthesized, it is sometimes necessary to distinguish parenthesized and unparenthesized calls in the mangling, despite the general rule that parentheses can be ignored. This is encoded using the choice of cl or cp for the call expression. The cp mangling is used only when the callee operand is a parenthesized unresolved name and would have used ADL if it were not parenthesized. In particular, cl is still used when unqualified lookup finds a declaration that would suppress the use of ADL, such as a class member.

5.1.6.1 Literals

Literal arguments, e.g. "A<42L>", are encoded with their type and value. Negative integer values are preceded with "n"; for example, "A<-42L>" becomes "1AILln42EE". The bool value false is encoded as 0, true as 1.

Floating-point literals are encoded using a fixed-length lowercase hexadecimal string corresponding to the internal representation, high-order bytes first. For example: "Lf bf800000 E" is -1.0f on platforms conforming to IEEE 754.

  <float> ::= <0-9a-f>+

The encoding for a literal of an enumerated type is the encoding of the type name followed by the encoding of the numeric value of the literal in its base integral type (which deals with values that don't have names declared in the type).

String literals are encoded using their type, but not their value. For example, L"abc" and L"123" are both encoded as "LA4_KwE" ("array [4] of const wchar_t").

The pointer literal expression nullptr is encoded as "LDnE". In contrast, a template argument which happens to be a null pointer (an extension made standard in C++11) is mangled as if it were a literal 0 of the appropriate pointer type; for example, "LPi0E" or "LDn0E". This inconsistency is an unfortunate accident.

5.1.6.2 References to declared entities

A reference to an entity with external linkage is encoded with "L<mangled name>E". For example:

          void foo(char); // mangled as _Z3fooc
          template<void (&)(char)> struct CB;
          // CB<foo> is mangled as "2CBIL_Z3foocEE"

The <encoding> of an extern "C" function is treated like global-scope data, i.e. as its <source-name> without a type. For example:

          extern "C" bool IsEmpty(char *); // (un)mangled as IsEmpty
          template<void (&)(char *)> struct CB;
          // CB<IsEmpty> is mangled as "2CBIL_Z7IsEmptyEE"

When encoding template signatures, a name appearing in the source code cannot always be resolved to a specific entity: In such cases the <encoding> production (via <expr-primary>) does not apply, and instead the <unresolved-name> encoding is used. For example:

          template<class T> auto f(T p)->decltype(p->x);
            // The return type in the mangling of the template signature
            // is encoded as "Dtptfp_1xE".
          template<class T> auto f(T p)->decltype(T::X::y);
            // The return type in the mangling of the template signature
            // is encoded as "DtsrNT_1XE1yE" (note how <type> is a
            // <nested-name> for T::X in this case).
          template<class T> auto f(T p)->decltype(p->::A::B::x);
            // The return type in the mangling of the template signature
            // is encoded as "Dtptfp_gssr1A1BE1xE".
          template<class T> auto f(T p)->decltype(p->x)::Y;
            // The return type in the mangling of the template signature
            // is encoded as "NDtptfp_1xE1YE".
In the case of member selection operations, the <unresolved-name> is used even if the indicated member is actually known. Similarly, an <unresolved-qualifier-level> may encode a known class type. That production is also used for references to nonstatic members with no associated expression designating the enclosing object (a C++11 feature). For example:
          struct Q { int x; } q;
          template<class T> auto f(T p)->decltype(p.x + q.x);
            // The return type in the mangling of the template signature
            // is encoded as "DTpldtfp_1xdtL_Z1qE1xE".
          template<class T> auto f(T p)->decltype(p.x + Q::x);
            // The return type in the mangling of the template signature
            // is encoded as "DTpldtfp_1xsr1QE1xE".
          template<class T> struct X { static T x; };
          struct B: X<int> {};
          struct D: B {} d;
          template<class T> auto f(T p)->decltype(p+d.B::X<T>::x);
            // The return type in the mangling of the template signature
            // is encoded as "DTplfp_dtL_Z1dEsr1B1XIT_EE1xE".  (The
            // "1B" part is a <unresolved-qualifier-level> encoding
            // a resolved type.)
If the <unresolved-name> refers to an operator for which both unary and binary manglings are available, the mangling chosen is the mangling for the binary version. For example:
          template<class T> auto f(T p)->decltype(&T::operator-);
            // The return type in the mangling of the template signature
            // is encoded as "DTadsrT_onmiE".

5.1.7 Scope Encoding

Entities declared in non-global scopes must include their scope in their mangled name. For entities declared outside of function definitions, this is dictated by the rules laid out for <name>. Entities declared within function definitions usually do not require a well-defined mangling because only one translation unit has access to the entity. However, different translation units must agree about the address of entities declared within inline functions, including template specializations. Therefore this ABI defines a mangling for even local entities.

The mangling of a local entity is composed of three elements: the mangling of the enclosing function, the mangling of the entity relative to the function, and an optional discriminator within the function:

  <local-name> ::= Z <function encoding> E <entity name> [<discriminator>]
               ::= Z <function encoding> E s [<discriminator>]

  <discriminator> ::= _ <non-negative number>      # when number < 10
                  ::= __ <non-negative number> _   # when number >= 10

The enclosing function is the closest function enclosing the entity. That is, when an entity E is declared within a function that is itself local to another function, the <encoding> beginning the mangling of E will itself be a <local-name>, perhaps recursively.

The name of a declared local entity is mangled with the rules for <name> as if the function were the global scope. That is, an entity declared directly within the function (e.g. a local type or static local variable) is mangled using an unscoped name, whereas a member of a local type will be mangled with a <nested-name>.

It is possible to declare multiple entities with the same name within a function if they are declared in different scopes. In this case, a discriminator must be added to the <local-name>. Entities with the same "top-level" name are numbered in lexical order within the function definition. A discriminator is added only for the second and later occurrences of the same name, and so the <number> in the discriminator is actually n-2 for the nth occurrence. "Top-level" here means the name declared directly in the local scope; e.g., if there are three classes named X in a given function g, and only the third has a member function f, the name of X::f in g will still include the discriminator _1 (because 3 - 2 == 1).

For unnamed local types (excluding types with a name for linkage purposes), the "name" is encoded as an <unnamed-type-name> of the form

 <unnamed-type-name> ::= Ut [<nonnegative number> ] _
where the number is is omitted for the first unnamed type in the function, and n-2 for the nth unnamed type (in lexical order) otherwise.

For example:

	inline void g(int) {
	  { struct S {}; }
	  { struct S {}; }
	  { struct S {}; }
	  struct S {        // Fourth occurrence: _2
	    void f(int) {   // _ZZ1giEN1S1fE_2i
	      struct {} x1;
	      struct {} x2;
	      struct {      // Third occurrence: 1_, i.e.
	                    // _ZZZ1giEN1S1fE_2iEUt1_
	        int fx() {  // _ZZZ1giEN1S1fE_2iENUt1_2fxEv
 	          return 3;
                }
	      } x3;
	      x3.fx();
	    }
	  } s;
	  s.f(1);
	}

The second production in <local-name> is used for string literals. The discriminator is used only if there is more than one, for the second and subsequent ones. In this case again <number> is n-2 for the nth distinct string literal, in lexical order, appearing in the function. Multiple references to the same string literal produce one string object with one name in the sequence. Note that this assumes that the same string literal occurring twice in a given function in fact represents a single entity, i.e. has a unique address.

For entities in constructors and destructors, the mangling of the complete object constructor or destructor is used as the base function name, i.e. the C1 or D1 version. This yields mangled names that are consistent across the versions.

Example:

	inline char const* g() {
	  "str1";                   // First string in g()
	  struct B {};
	  struct S: B {
	    S()                     // Complete object ctor: _ZZ1gvEN1SC1Ev
	      : msg("str2") {}      // First string in g()::S::S():
	                            //      _ZZZ1gvEN1SC1EvEs
	    char const *msg;
	  } s;
	  "str3";                   // Second string in g()
	  static char const *str4a  // _ZZ1gvE5str4a
	     = "str4";              // Third string in g() (n-2 == 1):
	                            //      _ZZ1gvEs_1
	  static char const *str4b  // _ZZ1gvE5str4b
	     = "str4";              // Still the third string (_ZZ1gvEs_1)
	  return str4b;
	}
See additional examples in the ABI examples document.

5.1.8 Closure Types (Lambdas)

A lambda expression introduces a unique class type called closure type. In some contexts, such closure types are unique to the translation unit: This ABI therefore does not specify an encoding for such cases (but an implementation must ensure that any internal encoding does not conflict with this ABI).

For example:

namespace N {
  int n = []{ return 1; }();  // Closure type internal to
}                             // the translation unit.
In the following contexts, however, the one-definition rule requires closure types in different translation units to "correspond": In all these contexts, the encoding of the closure types builds on an underlying <unqualified-name> that is an <unnamed-type-name> of the form
  <unnamed-type-name> ::= <closure-type-name>

  <closure-type-name> ::= Ul <lambda-sig> E [ <nonnegative number> ] _ 
with
  <lambda-sig> ::= <parameter type>+  # Parameter types or "v" if the lambda has no parameters
The number is omitted for the first closure type with a given <lambda-sig> in a given context; it is n-2 for the nth closure type (in lexical order) with that same <lambda-sig> and context.

If the context is the body of a function (inline and/or template), the closure type is encoded like any other local entity (see Scope Encoding above). For example:

	template<typename F> int algo(F fn) { return fn(); }
	inline void g(int n) {
	  int bef(int i = []{ return 1; }());
	    // Default arguments of block-extern function declarations
	    // remain in the context of the encloding function body.
	    // The closure type is encoded as Z1giEUlvE_.
	    // The call operator of that type is _ZZ1giENKUlvE_clEv.

	  algo([=]{return n+bef();});
	    // The captured entities do not participate in <lambda-sig>
	    // and so this closure type has the same <lambda-sig> as
	    // the previous one.  It encoding is therefore Z1giEUlvE0_
	    // and the call operator is _ZZ1giENKUlvE0_clEv.  The
	    // instance of "algo" being called is then
	    // _Z4algoIZ1giEUlvE0_EiT_.
	}

If the context is a default argument (of a member function parameter) appearing in a class definition, the closure class and its members are encoded as follows:

  <local-name> ::= Z <function encoding> Ed [ <parameter number> ] _ <entity name>
The parameter number is omitted for the last parameter, 0 for the second-to-last parameter, 1 for the third-to-last parameter, etc. The <entity name> will of course contain a <closure-type-name>: Its numbering will be local to the particular argument in which it appears -- other default arguments do not affect its encoding. For example:
	struct S {
	  void f(int = []{return 1;}()
	           // Type: ZN1S1fEiiEd0_UlvE_
	           // Operator: _ZZN1S1fEiiEd0_NKUlvE_clEv
	             + []{return 2;}(),
	           // Type: ZN1S1fEiiEd0_UlvE0_
	           // Operator: _ZZN1S1fEiiEd0_NKUlvE0_clEv
	         int = []{return 3;}());
	           // Type: ZN1S1fEiiEd_UlvE_
	           // Operator: _ZZN1S1fEiiEd_NKUlvE_clEv
	} s;

If the context of a closure type is an initializer for a class member (static or nonstatic), inline variable, or variable template, it is encoded in a qualified name with a <prefix> of the form:


    <closure-prefix> ::= [ <prefix> ] <variable or member unqualified-name> M
                     ::= <variable template template-prefix> <template-args> M
For example:
	template<typename T> struct S {
	  static int x;
	};
	template<typename T> int S<T>::x = []{return 1;}();
	template int S<int>::x;
	  // Type of lambda in intializer of S<int>::x: N1SIiE1xMUlvE_E
	  // Corresponding operator(): _ZNK1SIiE1xMUlvE_clEv

In a generic lambda, uses of auto in the parameter list are mangled as the corresponding artificial template type parameter. This is never ambiguous with a lambda parameter whose type is an enclosing template type parameter, because lambdas are never mangled in a dependent context (they are forbidden from appearing in function signatures). A <template-param> in a <lambda-sig> can only ever refer to a template parameter of a generic lambda.

5.1.9 Lexical ordering

Lexical ordering is used for numbering local entities (named and unnamed local classes and enumerations, closure types, and static local variables) when there is no other way of distinguishing them. Except as described below, all local entities are to be numbered, even if subsequent optimization makes some of them unnecessary, or no mangled name is actually required for some of them.

The order of entities is the source order of a key token unique to the entity as if:

In all cases, this is meant to imitate the numbering that would be produced by a simple implementation which numbered entities as it processed the source and recursively expanded packs. Since discarded substatements of if constexpr are still processed in non-template code, entities within them are still numbered.

Entities may be lexically nested without being in different contexts for the purposes of mangling. For example, the lambdas in [x = []{}]{} are both part of the enclosing context. The order of such entities is determined by the source order of a key token in the entity. In general, this token is the first token past which the signature of the entity (as is necessary to mangle it) is known:

All entities from one expansion of a pack are considered to occur lexically before any entities from the next expansion of the same pack, but the ordering is otherwise based on the original token sequence. This is expected to match the numbering that would be established for the corresponding non-template generated by substitution into the template. As a consequence of these rules, the entity instantiated for a particular source construct can have a different mangling number in different instantiations of the same template. An entity appearing in the original source may also go entirely un-numbered if it appears in a discarded sub-statement of if constexpr or in a pack expansion of an empty pack. For example:

void g(...);
template<bool b, typename ...T> void f() {
  if constexpr(b) { []{}; }
  g(([]{}, []{ static T n; return &n; }())...);
}

// The first lambda is discarded and does not receive a number.
// The variables passed to g are therefore mangled as
// _ZZZ1fILb0EJiiEEvvENKUlvE0_clEvE1n and
// _ZZZ1fILb0EJiiEEvvENKUlvE2_clEvE1n.
template void f<false, int, int>();

// The variable passed to g is mangled as
// _ZZZ1fILb1EJiEEvvENKUlvE1_clEvE1n and
template void f<true, int>();

// Both lambdas are numbered; returns _ZZZ1hvENKUlvE0_clEvE1n.
inline int *h() {
  if constexpr(false) { []{}; }
  return []{ static int n; return &n; }();
}

Note that the numbering of entities appearing within a mem-initializer-list reflects the order that the initializers appear in the source, which may be different from the order in which the initializers will be executed when the program runs.

5.1.10 Compression

To minimize the length of external names, we use two mechanisms, a substitution encoding to eliminate repetition of name components, and abbreviations for certain common names. Each non-terminal in the grammar above for which <substitution> appears on the right-hand side is both a source of future substitutions and a candidate for being substituted. There are two exceptions that appear to be substitution candidates from the grammar, but are explicitly excluded:

All substitutions are for entities that would appear in a symbol table. In particular, we make substitutions for prefixes of qualified names, but not for arbitrary components of them. Thus, the components ::n1::foo() and ::n2:foo() appearing in the same name would not result in substituting for the second "foo." Similarly, we do not substitute for expressions, though names appearing in them might be substituted. The reason for this is to facilitate implementations that use the symbol table to keep track of components that might be substitutable.

Note that the above exclusion of function and operator names from consideration for substitution does not exclude the full function entity, i.e. its name plus its signature encoding.

Logically, the substitutable components of a mangled name are considered left-to-right, components before the composite structure of which they are a part. If a component has been encountered before, it is substituted as described below. This decision is independent of whether its components have been substituted, so an implementation may optimize by considering large structures for substitution before their components. If a component has not been encountered before, its mangling is identified, and it is added to a dictionary of substitution candidates. No entity is added to the dictionary twice.

The type of a non-static member function is considered to be different, for the purposes of substitution, from the type of a namespace-scope or static member function whose type appears similar. The types of two non-static member functions are considered to be different, for the purposes of substitution, if the functions are members of different classes. In other words, for the purposes of substitution, the class of which the function is a member is considered part of the type of function.

<b>NOTE</b>: Therefore, in the following example:

typedef void T();
struct S {};
void f(T*, T (S::*)) {}
the function f is mangled as _Z1fPFvvEM1SFvvE; the type of the member function pointed to by the second parameter is not considered the same as the type of the function pointed to by the first parameter. Both function types are, however, entered the substitution table; subsequent references to either variant of the function type will result in the use of substitutions.

Substitutions are mangled using the usual rules for <seq-id>:


  <substitution> ::= S <seq-id> _
		 ::= S_

Substitutable components are numbered left-to-right. A component is earlier in the substitution dictionary than the structure of which it is a part. All substitutable components are numbered, except those that have already been numbered for substitution. For example:

   "_ZN1N1TIiiE2mfES0_IddE": Ret? N::T<int, int>::mf(N::T<double, double>)
since the substitutions generated for this name are:
   "S_" == N (qualifier is less recent than qualified entity)
   "S0_" == N::T (template-id comes before template)
	(int is builtin, and isn't considered)
   "S1_" == N::T<int, int>
   "S2_" == N::T<double, double>

Note that substitutable components are the represented symbolic constructs, not their associated mangling character strings. Thus, a substituted object matches its unsubstituted form, and a delimited <function-type> matches its <bare-function-type>.

In addition, the following catalog of abbreviations of the form "Sx" are used:


   <substitution> ::= St # ::std::
   <substitution> ::= Sa # ::std::allocator
   <substitution> ::= Sb # ::std::basic_string
   <substitution> ::= Ss # ::std::basic_string < char,
						 ::std::char_traits<char>,
						 ::std::allocator<char> >
   <substitution> ::= Si # ::std::basic_istream<char,  std::char_traits<char> >
   <substitution> ::= So # ::std::basic_ostream<char,  std::char_traits<char> >
   <substitution> ::= Sd # ::std::basic_iostream<char, std::char_traits<char> >

Note that the abbreviation St does not require N...E delimiters unless either followed by more than one additional composite name component, or preceded by CV-qualifiers or a ref-qualifier for a member function. For example:

   "_ZSt5state": ::std::state
   "_ZNSt3_In4wardE": ::std::_In::ward


5.2 Vague Linkage

Many objects in C++ are not clearly part of a single object file, but are required by the ODR to have a single definition. This section identifies, for such objects, where (i.e. in which objects) they should be emitted, and what special treatment might be required if duplicates are possible.

In many cases, we will deal with duplicates by putting possibly duplicated objects in distinct ELF sections or groups of sections, and using the COMDAT feature of SHT_GROUP sections in the gABI to remove duplicates. We will refer to this simply as using a COMDAT group, and specify the symbol to be used to identify duplicates in the SHT_GROUP section. COMDAT groups are a new gABI feature specified during the Itanium ABI definition, and may not be implemented everywhere immediately. See the separate ABI examples document for a discussion of alternatives pending COMDAT implementation.

Note that nothing in this section should be construed to require COMDAT usage for objects with internal linkage unless they may in fact be referenced outside the translation unit where they appear, for instance due to inlining.

5.2.1 Out-of-line Functions

It may sometimes be necessary or desirable to reference an out-of-line copy of a function declared inline, i.e. to reference a global symbol naming the function. This may occur because the implementation cannot, or chooses not to, inline the function, or because it needs an address rather than a call. In such a case, the function is to be emitted in each object where its name is referenced. A COMDAT group is used to eliminate duplicates, with the mangled name of the function as the identifying symbol.

5.2.2 Static Data

Inline functions, whether or not declared as such, and whether they are inline or out-of-line copies, may reference static data or character string literals, that must be kept in common among all copies by using the local symbol mangling defined above. These objects are named according to the rules for local names in the Scope Encoding section above, and the definition of each is emitted in a COMDAT group, identified by the symbol name described in the Scope Encoding section above. Each COMDAT group must be emitted in any object with references to the symbol for the object it contains, whether inline or out-of-line.

Some objects with static storage duration have associated guard variables used to ensure that they are initialized only once (see 3.3.3). If the object is emitted using a COMDAT group, the guard variable must be too. It is suggested that it be emitted in the same COMDAT group as the associated data object, but it may be emitted in its own COMDAT group, identified by its name. In either case, it must be weak.

5.2.3 Virtual Tables

Virtual tables for dynamic classes are emitted as follows:

All of the following must be emitted whenever tables are emitted:

If the emission of the virtual tables is not unique (as defined above), they are emitted with vague linkage in a single COMDAT group, using the mangled name for the virtual table as the COMDAT identifier. On platforms that do not use COMDATs, the tables are emitted separately, and unreferenced tables can be omitted.

Note: previous versions of this ABI required virtual tables to always be emitted with vague linkage, even when emitted in a unique object. This is unnecessary and should not detectable in a valid program; however, it is also harmless, and implementations may opt to continue to do it for compatibility with programs that are not strictly valid.
Note: the requirement to emit all of the tables for a class in a single COMDAT does not currently serve any clear purpose beyond reducing the number of COMDATs. There is no known optimization that would result in e.g. a VTT that would only work correctly with the paired virtual table or vice-versa. Nonetheless, it is required by the ABI and is largely harmless.

The key function is the first non-pure virtual function that is not inline at the point of class definition. constexpr or consteval functions are always declared as such on their first declarations and so are implicitly inline, so they are never key functions.

Note that if the key function is not inline in the class definition, but its later definition is inline, it will be emitted in every object containing the definition.

Recommendation for new platforms: when selecting the key function, ignore virtual function definitions that are inline outside of the class definition. Inline function definitions must be defined in every translation unit that ODR-use them, and virtual functions are considered to always be ODR-used. Therefore, in a valid program, every translation unit that sees a class definition must agree about which of its virtual functions are defined inline outside of the class. Some platforms are known to have already adopted this rule.
Recommendation for new platforms: when selecting the key function, also consider pure virtual destructors as candidates. Virtual destructors must be defined even if they are pure.

5.2.4 std::type_info

The RTTI std::type_info structure for a complete dynamic class type whose virtual tables are emitted in a unique object (see above) is emitted with non-vague linkage in the same object as its virtual tables.

The RTTI std::type_info structure for an incomplete class type is emitted wherever referenced. It is implementation-defined what symbol name it uses, other than that it must not be the same symbol name as any other entity, including the complete class type. It also implementation-defined whether they are emitted in a COMDAT group or just with internal linkage.

The RTTI std::type_info structures for various basic types as specified by the Run-Time Type Information section are provided by the runtime library.

The RTTI std::type_info structure for any other type is emitted with vague linkage using the symbol name as the COMDAT identifier.

The RTTI NTBS name object for a type is emitted in every object that references it, with vague linkage using the NTBS mangled name as the COMDAT identifier. Vague linkage is required even for classes where the std::type_info object has non-vague linkage in order to support interoperation with incomplete RTTI for incomplete types.

Recommendation for new platforms: the linkage rules for RTTI std::type_info and name objects are known to cause both semantic and performance problems when dynamic linking is in use. This is an area where there is widespread divergence from this ABI, chiefly to avoid the need for cross-image symbol coalescing. Several different approaches for this are in use, with different underlying goals. For now, there is not a clear recommendation to make, other than that the current rule is not the best.

5.2.5 Constructors and Destructors

Constructors and destructors for a class, whether implicitly-defined or user-defined, are emitted under the same rules as other functions. That is, user-defined constructors or destructors, unless the function is declared inline, or has internal linkage, are emitted where defined, with their complete, and base object variants. For destructors, in classes with a virtual destructor, the deleting variant is emitted as well. A user-defined constructor or destructor with non-inline, internal linkage is emitted where defined, with only the variants actually referenced. Implicitly-defined or inline user-defined constructors and destructors are emitted where referenced, each in its own COMDAT group identified by the constructor or destructor name.

This ABI does not require the generation or use of allocating constructors or inheriting constructors, and does not require the generation or use of deleting destructors for classes without a virtual destructor. However, if an implementation emits such functions, it must use the external names specified in this ABI. If such a function has external linkage, it must be emitted wherever referenced, in a COMDAT group whose name is the external name of the function.

5.2.6 Instantiated Templates

An instantiation of a class template requires:

An instantiation of a function template or member function template is emitted in any object where its symbol is referenced (non-inline), in a COMDAT group identified by the function name.


5.3 Unwind Table Location

As described in the Itanium psABI, Itanium implementations shall produce unwind table entries in a SHT_IA_64_UNWIND section, and unwind information descriptors in a section that will be linked with the associated code. Itanium linkers shall put the unwind table, the unwind information table, and the associated code in a single text segment, with a PT_IA_64_UNWIND program table entry identifying the unwind table location.



Appendix R: Revision History


[160602] Describe abi_tag mangling.

[160907] Introduce layout restrictions on virtual table pointers stored in objects and VTTs.

[151021] Support transaction-safe functions.

[151019] Add mangling for unresolved names rooted in template template parameters.

[150518] Allow arbitrary arguments (encoded as template arguments) in the mangling of vendor-specific type qualifiers.

[150502] Clarify mangling of nested and local names and correct an ambiguity in prefix mangling.

[150204] Fix alignment calculation for empty proper base classes.

[150204] Add mangling for lifetime-extended temporaries.

[150204] Add mangling for captured template parameter packs.

[150204] Add mangling for braced initializer lists.

[150204] Define behavior for variadic arguments of non-trivial type.

[140427] Add mangling for dependent elaborated type specifiers.

[130911] Add mangling for null template arguments.

[130710] Add mangling for operator "".

[130617] Fix an editorial error in the mangling of floating-point literals.

[130606] Clarify rules for POD types in the face of the C++11 changes to the definition of POD. Minor restructuring for clarity in the mangling section.

[130422] Add mangling for decltype(auto).

[130403] Add mangling for ref-qualifiers on function types.

[121211] Add __cxa_throw_bad_array_new_length mangling.

[120925] Add noexcept mangling.

[110306] Update description of mangling for argument packs.

[110306] Update description of mangling for argument packs.

[110301] Change mangling for argument packs.

[101124] Revise mangling specification to cover instantiation-dependent expressions.

[100625] Add nullptr_t mangling.

[100212] Permit mangling of additional expression forms as template arguments.

[091124] Document passing for IEEE 754r decimal and half-precision floating point types.

[091113] Document mangling for IEEE 754r decimal and half-precision floating point types.

[091007] Document handling of lambdas.

[090715] Document handling of deleted virtual functions.

[090312] Remove type stub expressions. Add mangling for alignof, function parameters, and a different mangling for N-argument function casts.

[090102] Remove mangling for N-argument functional casts.

[081210] Add manglings for type stub expressions, call expressions, char*_t, and N-argument functional casts. Change argument pack mangling.

[080707] Add manglings for IEEE 754r decimal and half-precision floating point types.

[072507] Add mangling for variadic templates and decltype.

[071207] Add mangling for rvalue references.

[031006] Clarify that guard variables are used to guard static data members of class templates, as well as function-scope statics.

[030806] Specify that function pointers in virtual tables are address/GP pairs on Itanium.

[050504] Remove use of out0 for by-value return types on Itanium.

[050211] Reverse treatment of ambiguous arguments to __cxa_demangle (3.4).

[041118] Clarify the layout of bit-fields.

[041025] Indicate that the TC1 definition of POD is intended in the section defining a "POD for the purpose of layout". Clearly indicate that an array whose elements are not PODs for the purpose of layout is itself not a POD for the purpose of layout.

[040923] Clarify behavior of __cxa_vec_delete.

[040219] Clarify substition of member function types.

[031128] Fix alphabetization of company names.

[031123] Add note about forward references to template parameters in member template conversion operators.

[031102] Specify the behavior of __cxa_vec_delete when the array_address is NULL.

[030905] Specify the behavior of __cxa_vec_new, __cxa_vec_new2, and __cxa_vec_new3 in the event that the allocation function returns NULL.

[030609] Use void* instead of dso_handle.

[030518] Specify behavior of __cxa_vec_new2 and __cxa_vec_new3 when the deallocation function throws an exception.

[030518] Define "POD for the purpose of layout."

[030316] Add acknowledgements section.

[030313] Correct broken links and incorrect formatting.

[030103] Clarify definition of substantively different types.

[021222] Document mangling for anonymous unions.

[021204] Remove note about 32-bit RTTI variation.

[021125] Clarify guard functions.

[021110] Clarify definition of nearly empty class.

[021110] Clarify ordering of string literals in mem-initializer-list.

[021110] Remove unnecessary V-adjusting thunks.

[021110] Clarify VTT contents.

[021021] Specify place and manner of emission for deleting destructors.

[021021] Clarify mangling of pointer-to-member functions.

[021016] Clarify mangling of floating-point literals.

[021014] Clarify use of sr in mangling.

[021011] Add mangling for unary plus.

[021008] Make the names used for constructors and destructor entry points consistent throughout.

[021008] Define manglings for typename types.

[020916] Clarify ordering of functions in virtual function table. Correct mangling substitution example.

[020906] Add ternary expression variant. Remove use of "low-order" to describe bytes in guard variables.

[020827] Clarify definition of nearly empty class, dsize, nvsize, nvalign.

[020827] Clarify handling of tail-padding.

[020326] Clarify wording in __cxa_demangle memory management specification.

[020220] Clarify pointer to member function mangling (5.1.5).

[010407] Don't assume that virtual functions can be called through intermediate bases. Add notes about missed opportunities. The VTT parm isn't mangled, either.

[010315] Many outstanding updates. Empty classes passed as ordinary classes (3.1.3). Secondary virtual pointers for subobjects reachable via a virtual path (text of 2.6.1, text and example in 2.6.2). Note about locating virtual bases statically during construction (2.6.1). Rename IA-64 to Itanium throughout. Add __cxa_vec_cleanup (3.3.4).

[000817] Updates from 17 August meeting, email.

[000807] Added base document section (1.5). Further RTTI field name cleanup (2.9.4). Update proposed one-time construction API (3.3.3). Update proposed object construction priority API (3.3.5). Removed <name> substitution (5.1.2). COMDAT not generally necessary for internal linkage (5.2). COMDAT for local static guard variables (5.2.2).

[000727] Updates from 20 July meeting. Added section on controlling object construction order (3.3.5).

[000707] Introduce consistent type_info field names (2.9.4). Removed vmi flags for publicly/non-publicly inherited bases (2.9.4). Collect all construction/destruction APIs in one section (3.3). Added one-time initialization API (3.3.3). Vector construction/destruction routines are extern "C" (3.3.4). Added routines for vector construction/destruction (3.3.4). Added copy construction runtime API (3.3.4). Make Alex's changes in mangling grammar (5.1). Add <special-name> cases for covariant override thunks (5.1.4). Allow expressions as array type dimensions (5.1.5). Discuss vague linkage for virtual function override thunks (5.2.6).

[000621] Add scope section 1.4. Specify guard variables and vague linkage of static data (5.2.2) and instantiated templates (5.2.4). Clarify vcall offsets (2.5.3), VTT (2.6.2), mangling compression rules (5.1.7), and mangling examples.

[000511] Specify 32-bit form of vmi_offset_flags. Add export template note.

[000505] Updates from 4 May meeting. VTT is preorder, like everything else. Add issue C-3 destructor API. Added demangler API. Yet another try at the nested-name mangling grammar. Don't mangle builtin types (except vendor extended ones). Reverse mangling substitution order, and fix mangling substitution examples. Add vague linkage information for instantiated templates. Specify location of unwind tables.

[000502] Fixed mangling of template parameters again.

[000427] Reorganization and section numbering. Added non-virtual function calling conventions.

[000417] Updates from 17 April meeting. Clarify order of vcall offsets. More elaboration of construction virtual table. Specification of COMDAT RTTI name. Reorganization of pointer RTTI. Modify mangling grammar to clarify substitution in compound names. Clarify Vague Linkage section.

[000407] Updates from 6 April meeting, email. More elaboration of construction vtable. Updates/issues in RTTI. Minor mangling changes. Added Vague Linkage section.

[000327] Updates from 30 March meeting. Define base classes to include self, proper base classes. Modify local function mangling per JFW proposal.

[000327] Updates from 23 March meeting. Adopt construction vtable Proposal B, and rewrite. Further work on mangling, especially substitution.

[000320] Clarify class size limit. Editorial changes in vtable components description. Add alternate to construction vtable proposal. Clarification in array cookie specification. Removed COMMON proxy from class RTTI. Extensive changes to mangling writeup.

[000314] Construction vtable modifications. RTTI modifications for incomplete class types. Mangling rework: grammar, new constructs, function return types.

[000309] Add limits section. Specify NULL member pointer values. Combine vtable content and order sections; clarify ordering. Specify when distinct virtual function entries are needed for overriders. Define (and modify) vector constructor/destructor runtime APIs. Virtual base offsets are promoted from non-virtual bases.

[000228] Add thunk definition. Revise inheritance graph order definition. Fix member function pointer description (no division by two). Move bit-field allocation description (much modified) to the non-virtual-base allocation description. Replace virtual function calling convention description.

[000228] Add thunk definition. Revise inheritance graph order definition. Fix member function pointer description (no division by two). Move bit-field allocation description (much modified) to the non-virtual-base allocation description. Replace virtual function calling convention description.

[000217] Add excess-size bit-field specification. Add namespace/header section. Touch up array new cookies. Remove construction vtable example to new file. Add mangling proposal.

[000214] Complete array new cookie specification. Remove unnecessary RTTI flags. Correct repeated inheritance flag description. Move all type_info subclasses in namespace abi, not namespace std. Note requirements for an implementation to prevent users from emitting invalid vtables for RTTI classes. Include construction vtable proposal.

[000203] Incorporate discussion of 3 Febrary. Remove __reference_type_info (issue A-22). Restructure struct RTTI and flags (issue A-23). Clarify __base_class_info layout.

[000125] Incorporate discussion of 20 January, generally clarifications. Resolved A-19 (choice of a primary virtual base). Answered Nathan's questions about RTTI. Included RTTI "Deliberations" as rationale notes in the specification, or removed redundant ones. Added array operator new section.

[000119] Clarify when virtual base offsets are required. Note that a vtable has offset-to-top and RTTI entries for classes with virtual bases even if there are no virtual functions. Resolve allocation of a virtual base class that is a primary base for another base (A-17). Resolve choice of a primary virtual base class that is a primary base for another base (A-19). Describe the (non-)effect of virtual bases on the alignment of the non-virtual part of a class as the base of another class (A-18).

[991230] Integrate proposed resolution of A-16, A-17 in base class layout. Add outstanding questions list, and clean up questions in text.

[991229] Clarify definition of nearly empty class, layout of virtual bases.

font color=blue>[991203] Added description of vfunc calling convention from Jason.

[991104] Noted pair of vtable entries for virtual destructors.

[991019] Modified RTTI proposal for 14 October decisions.

[991006] Added RTTI proposal.

[990930] Updated to new vtable layout proposal.

[990811] Described member pointer representations, virtual table layout.

[990730] Selected first variant for empty base allocation; removed others.