Itanium C++ ABI ($Revision: 1.86 $)


Contents


Acknowledgements


This document was 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.


Chapter 1: Introduction


In this document, we specify the Application Binary Interface for C++ programs, that is, the object code interfaces between user C++ code and the implementation-provided system 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 written as a generic specification, to be usable by C++ implementations on a variety of architectures. However, it does contain processor-specific material for the Itanium 64-bit ABI, identified as such. Where structured data layout is described, we generally assume Itanium psABI member sizes. An implementation for a 32-bit ABI would typically just change the sizes of members as appropriate (i.e. pointers and long ints would become 32 bits), but sometimes an order change would be required for compactness, and we note more substantive changes.


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.

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 zero-width bitfields, no virtual functions, no virtual base classes, and no non-empty non-virtual proper base classes.

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.

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.

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 POD-struct or POD-union (in the sense of ISO C++ [class]) with a bitfield member whose declared width is wider than the declared type of the bitfield is not a POD for the purpose of layout. Similarly, an array type is not a POD for the purpose of layout if the element type of the array is not a POD for the purpose of layout. Where references to the ISO C++ are made in this paragraph, the Technical Corrigendum 1 version of the standard is intended.

<b>NOTE</b>: The ISO C++ standard published in 1998 had a different definition of POD types. In particular, a class with a non-static data member of pointer-to-member type was not considered a POD in C++98, but is considered a POD in TC1. Because the C++ standard requires that compilers not overlay the tail padding in a POD, using the C++98 definition in this ABI would prevent a conforming compiler from correctly implementing the TC1 version of the C++ standard. Therefore, this ABI uses the TC1 definition of POD.

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.

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. Type bool has size and alignment 1. All of these types have data size and non-virtual size equal to their size. (We ignore tail padding for PODs because the Standard does not allow us to use it for anything else.)


2.3 Member Pointers

A pointer to data member is an offset from the base address of the class object containing it, represented as a ptrdiff_t. It has the size and alignment attributes of a ptrdiff_t. A NULL pointer is represented as -1.

A pointer to member function is a pair as follows:

ptr:
For a non-virtual function, this field is a simple function pointer. (Under current base Itanium psABI conventions, that is a pointer to a GP/function address pair.) For a virtual function, it is 1 plus the virtual table offset (in bytes) of the function, represented as a ptrdiff_t. The value zero represents a NULL pointer, independent of the adjustment field value below.

adj:
The required adjustment to this, represented as a ptrdiff_t.

It has the size, data size, and alignment of a class containing those two members, in that order. (For 64-bit Itanium, that will be 16, 16, and 8 bytes respectively.)


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.) Further, assume for data members that nvsize==size, and nvalign==align. 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 bitfields in declaration order), allocate as follows:

    1. If D is a (possibly unnamed) bitfield 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 bitfield is allocated as required by the underlying C psABI, subject to the constraint that a bitfield 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 bitfield, and that bitfield 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 bitfield 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 bitfield, 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 bitfield, and update sizeof(C) to max(sizeof(C),dsize(C)).

    2. If D is not an empty base class or D is a 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 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 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:

      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. 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)). Note that nvalign(D) is 1, so no update of align(C) is needed. Similarly, since D is an empty base class, 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

    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 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:

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. There is one function pointer entry for each virtual function declared in the class, in declaration order, with any implicitly-defined virtual destructor pair last.

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 declared in the derived class (in declaration order) for which one of the following two conditions holds:

<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 following function pointer entries, following those of any primary base's virtual table, in the virtual functions' declaration order:

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 of the first virtual function pointer in the virtual table, 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 subclass 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 subclass, 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, 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 variables 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 3.3.2 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, 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, char16_t, char32_t, and the IEEE 754r decimal and half-precision floating point types. (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 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.

2.9.5 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
      	  };
            };
      
      

      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
    	  };
          };
    
    

    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:

    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;
          };
    
    

<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.

2.9.6 std::type_info::name()

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

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: Function Calling Conventions and APIs


In general, the calling conventions for C++ in this ABI follow those specified by the underlying processor-specific ABI for C, whenever there is an analogous construct in C. This chapter specifies exceptions required by C++-specific semantics, or by features without analogues in C. 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 Non-Virtual Function Calling Conventions

3.1.1 Value Parameters

In general, C++ value parameters are handled just like C parameters. This includes class type parameters passed wholly or partially in registers. There are, however, some special cases.

  1. In the special case where the parameter type has a non-trivial copy constructor or destructor, the caller must allocate space for a temporary copy, and pass the resulting copy by reference (below). Specifically,

  2. In the case where the parameter type is class std::decimal::decimal32, std::decimal::decimal64, or std::decimal::decimal128 as defined in TR 24733, the parameter is passed the same as the corresponding native decimal floating-point scalar type.

3.1.2 Reference Parameters

Reference parameters are handled by passing a pointer to the actual parameter.

3.1.3 Empty Parameters

Empty classes will be passed no differently from ordinary classes. If passed in registers the NaT bit must not be set on all registers that make up the class.

The contents of the single byte parameter slot are unspecified, and the callee may not depend on any particular value. On Itanium, the associated NaT bit must not be set if the parameter slot is associated with a register.

3.1.4 Return Values

In general, C++ return values are handled just like C return values. This includes class type results returned in registers. However, if the return value type has a non-trivial copy constructor or destructor, the caller allocates space for a temporary, and passes a pointer to the temporary as an implicit first parameter preceding both the this parameter and user parameters. The callee constructs the return value into this temporary.

Another exception is that a return value type 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.

A result of an empty class type will be returned as though it were a struct containing a single char, i.e. struct S { char c; };. The actual content of the return register is unspecified. On Itanium, the associated NaT bit must not be set.

3.1.5 Constructor Return Values

Constructors return void results.

3.2 Virtual Function Calling Conventions

3.2.1 Foundation

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 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.

3.3.3 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.

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_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.

3.3.4 Controlling Object Construction Order

3.3.4.1 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.

3.3.4.2 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.

3.3.4.3 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.

3.3.4.4 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.

3.3.4.5 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.

3.3.5 DSO Object Destruction API

3.3.5.1 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).

3.3.5.2 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.

3.3.5.3 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.

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>
    <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). The type is omitted for variables and static data members.

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.

    <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>
Names of objects nested in namespaces or classes are identified as a delimited sequence of names identifying the enclosing scopes. In addition, when naming a class member function, CV-qualifiers may be prefixed to the compound name, encoding the this attributes. Note that if member function CV-qualifiers are required, the delimited form must be used even if the remainder of the name is a single substitution.

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

    <prefix> ::= <prefix> <unqualified-name>
	     ::= <template-prefix> <template-args>
             ::= <template-param>
	     ::= # empty
	     ::= <substitution>
             ::= <prefix> <data-member-prefix>

    <template-prefix> ::= <prefix> <template unqualified-name>
                      ::= <template-param>
                      ::= <substitution>
    <unqualified-name> ::= <operator-name>
                       ::= <ctor-dtor-name>  
                       ::= <source-name>   
                       ::= <unnamed-type-name>   

    <source-name> ::= <positive length number> <identifier>
    <number> ::= [n] <non-negative decimal integer>
    <identifier> ::= <unqualified source code identifier>

<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'. <identifier> is a pseudo-terminal representing the unqualified identifier for the entity in the source code.

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>.

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. All operators are encoded using exactly two letters, the first of which is lowercase.


  <operator-name> ::= nw	# new           
		  ::= na	# new[]
		  ::= dl	# delete        
		  ::= da	# delete[]      
		  ::= 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	# >=            
		  ::= nt	# !             
		  ::= aa	# &&            
		  ::= oo	# ||            
		  ::= pp	# ++            
		  ::= mm	# --            
		  ::= cm	# ,             
		  ::= pm	# ->*           
		  ::= pt	# ->            
		  ::= cl	# ()            
		  ::= ix	# []            
		  ::= qu	# ?             
		  ::= st	# sizeof (a type)
		  ::= sz	# sizeof (an expression)
                  ::= at        # alignof (a type)
                  ::= az        # alignof (an expression)
		  ::= cv <type>	# (cast)        
		  ::= 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

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)

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>

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

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
		   ::= D0	# deleting destructor
		   ::= D1	# complete object destructor
		   ::= D2	# base object destructor

5.1.5 Type encodings

Types are encoded as follows:


  <type> ::= <builtin-type>
	 ::= <function-type>
	 ::= <class-enum-type>
	 ::= <array-type>
	 ::= <pointer-to-member-type>
	 ::= <template-param>
	 ::= <template-template-param> <template-args>
	 ::= <substitution> # See Compression below

Types are qualified (optionally) by single-character prefixes encoding cv-qualifiers and/or pointer, reference, complex, or imaginary types:


  <type> ::= <CV-qualifiers> <type>
	 ::= P <type>	# pointer-to
	 ::= R <type>	# reference-to
	 ::= O <type>	# rvalue reference-to (C++0x)
	 ::= C <type>	# complex pair (C 2000)
	 ::= G <type>	# imaginary (C 2000)
	 ::= U <source-name> <type>	# vendor extended type qualifier

  <CV-qualifiers> ::= [r] [V] [K] 	# restrict (C99), volatile, const

Vendors who define extended type qualifiers (e.g. _near, _far for pointers) shall encode them as a 'U' prefix followed by the name in <length,ID> form.

In cases where multiple order-insensitive qualifiers are present, they should be ordered 'K' (closest to the base type), '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 p has mangled type name U4_farrVKPi.

Vendors must therefore specify which of their extended qualifiers are considered order-insensitive, not necessarily 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 CV-qualified type, the base type is substitutible, and the type with all the C, V, and r qualifiers plus any vendor extended types in the same order-insensitive set is substitutible; any type with a subset of those qualifiers is 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. Also, note that the grammar above is written with the assumption that vendor extended type qualifiers will be in the order-sensitive (not CV) set. An appropriate grammar modification would be necessitated by an order-insensitive vendor extended type qualifier like const or volatile.

<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.

C++0x pack expansions are prefixed with Dp. The C++0x decltype type is prefixed with either Dt or DT, depending on how the decltype type was parsed.


 <type>  ::= Dp <type>          # pack expansion of (C++0x)
         ::= Dt <expression> E  # decltype of an id-expression or class member access (C++0x)
         ::= DT <expression> E  # decltype of an expression (C++0x)

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)
                 ::= Di # char32_t
                 ::= Ds # char16_t
		 ::= u <source-name>	# vendor extended type

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

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-parameter> 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.

A "Y" prefix for the bare function type encodes extern "C". If there are any cv-qualifiers of this, they are encoded at the beginning of the <qualified-name> as described above. This affects only type mangling, since extern "C" function objects have unmangled names.


  <function-type> ::= F [Y] <bare-function-type> E
  <bare-function-type> ::= <signature type>+
	# types are possible return type, then parameter types

When a function parameter is a C++0x function parameter pack, its type is mangled with Dp <type>, i.e., its type is a pack expansion.

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>

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

Array types encode the dimension (number of elements) and the element type. Note that "array" parameters to functions are encoded as pointer types. For variable length arrays (C99 VLAs), the dimension (but not the '_' separator) is omitted.


  <array-type> ::= A <positive dimension number> _ <element type>
	       ::= A [<dimension expression>] _ <element type>

When the dimension is an expression involving template parameters, the second production is used. Thus, the declarations:

    template<int I> void foo (int (&)[I + 1]) { }
    template void foo<2> (int (&)[3]);
produce the mangled name "_Z3fooILi2EEvRAplT_Li1E_i".

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


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

Note that for a pointer to cv-qualified member function, the qualifiers are attached to the function type, so


    struct A;
    void f (void (A::*)() const) {}
produces the mangled name "_Z1fM1AKFvvE".

When function and member function template instantiations reference the template parameters in their parameter/result types, the template parameter number is encoded, with the sequence T_, T0_, ... Class template parameter references are mangled using the standard mangling for the actual parameter type, typically a substitution. Note that a template parameter reference is a substitution candidate, distinct from the type (or other substitutible entity) that is the actual parameter.


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

Function parameters referenced in a late-specified return type are handled similarly:


  <function-param> ::= fp_	# first function parameter
		   ::= fp <parameter-2 non-negative number> _

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 additional I/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
                 ::= I <template-arg>* E                           # argument pack
                 ::= sp <expression>                               # pack expansion of (C++0x)

  <expression> ::= <unary operator-name> <expression>
               ::= <binary operator-name> <expression> <expression>
               ::= <trinary operator-name> <expression> <expression> <expression>
               ::= cl <expression>+ E                              # call
               ::= cv <type> <expression>                          # conversion with one argument
               ::= cv <type> _ <expression>* E                     # conversion with a different number of arguments
               ::= st <type>                                       # sizeof (a type)
               ::= at <type>                                       # alignof (a type)
               ::= <template-param>
               ::= <function-param>
               ::= sr <type> <unqualified-name>                    # dependent name
               ::= sr <type> <unqualified-name> <template-args>    # dependent template-id
               ::= dt <expression> <unqualified-name>              # expr.name
               ::= dt <expression> <unqualified-name> <template-args>
               ::= pt <expression> <unqualified-name>              # expr->name
               ::= pt <expression> <unqualified-name> <template-args>
               ::= on <operator-name>                              # dependent operator-function-id
               ::= on <operator-name> <template-args>              # dependent operator template-id
               ::= <source-name>                                   # dependent name
               ::= <source-name> <template-args>                   # dependent template-id
               ::= sZ <template-param>                             # size of a parameter pack
               ::= <expr-primary>

  <expr-primary> ::= L <type> <value number> E                     # integer literal
                 ::= L <type> <value float> E                      # floating literal
                 ::= L <mangled-name> E                            # external name

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.

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 (IEEE on Itanium), high-order bytes first, without leading zeroes. For example: "Lf bf800000 E" is -1.0f on Itanium.

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).

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"

An expression, e.g., "B<(J+1)/2>", is encoded with a prefix traversal of the operators involved. The operators are encoded using their two letter mangled names. For example, "B<(J+1)/2>", if J is the third template parameter, becomes "1BI Xdv pl T1_ Li1E Li2E E E" (the blanks are present only to visualize the decomposition). Note that the expression is mangled without constant folding or other simplification, and without parentheses, which are implicit in the prefix representation. Except for the parentheses, therefore, it represents the source token stream. (C++ Standard reference 14.5.5.1 p. 5.) An expression used as a template argument is delimited by "X...E".

If an expression is a qualified-name, and the qualifying scope is a dependent type, one of the sr productions is used, rather than the <mangled-name> production. If the qualified name refers to an operator for which both unary and binary manglings are available, the mangling chosen is the mangling for the binary version.

5.1.6 Scope Encoding

A nonlocal scope is encoded as the qualifier of a qualified name: it can be the top-level name qualification or it can appear inside <type> to denote dependent types or bind specific names as arguments. Qualified names are encoded as:

   N <qual 1> ... <qual N> <unqual name> E
where each <qual K> is the encoding of a namespace name or a class name (with the latter possibly including a template argument list).

Occasionally entities in local scopes must be mangled too (e.g. because inlining or template compilation causes multiple translation units to require access to that entity). The encoding for such entities is as follows:

  <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 first production is used for named local static objects and classes, which are identified by their "names" as encoded relative to the closest enclosing function. In case of unnamed local types (excluding unnamed types that have acquired a "name for linkage purposes"), the "name" the unqualified 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. The <entity name> may itself be a compound name, but it is relative to the closest enclosing function, i.e. none of the components of the function encoding appear in the entity name. It is possible to have nested function scopes, e.g. when dealing with a member function in a local class. In such cases, the function encoding will itself have <local-name> structure.

The discriminator is used only for the second and later occurrences of the same "top-level" name within a single function (since "unnamed types" are distinctly numbered, they never include a discriminator). In this case <number> is n - 2, if this is the nth occurrence, in lexical order, of the given name. "top-level" here means that if there are e.g. three classes named X in a given function g, and only the third has a member function f, the encoding of S::f in g will still include a discriminator of the form "_1" (n-2 == 1).

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 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 <number> is n - 2, if this is 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.

In all cases the numbering order is strictly lexical order based on the original token sequence. All entities occurring in that sequence are to be numbered, even if subsequent optimization makes some of them unnecessary. The ordering of literals appearing in a mem-initializer-list shall be the order that the literals appear in the source, which may be different from the order in which the initializers will be executed when the program runs. It is expected that this will be the 'natural' order in most compilers. In any case, conflicts would arise only if different compilation units including the same code were compiled by different compilers, and multiple entities requiring mangling had the same name.

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.

5.1.7 Closure Types (Lambdas)

A C++0x 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;

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

  <data-member-prefix> := <member source-name> 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

5.1.8 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.

Substitution is according to the production:


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

The <seq-id> is a sequence number in base 36, using digits and upper case letters, and identifies the <seq-id>-th encoded component, in left-to-right order, starting at "0". As a special case, the first substitutable entity is encoded as "S_", i.e. with no number, so the numbered entities are the second one as "S0_", the third as "S1_", the twelfth as "SA_", the thirty-eighth as "S10_", etc. All substitutable components are so numbered, except those that have already been numbered for substitution. A component is earlier in the substitution dictionary than the structure of which it is a part. 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> >

The abbreviation St is always an initial qualifier, i.e. appearing as the first element of a compound name. It does not require N...E delimiters unless either followed by more than one additional composite name component, or preceded by CV-qualifiers for a member function. This adds the case:


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

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 3.3.2). 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

The virtual table for a class is emitted in the same object containing the definition of its key function, i.e. the first non-pure virtual function that is not inline at the point of class definition. If there is no key function, it is emitted everywhere used. The emitted virtual table includes the full virtual table group for the class, any new construction virtual tables required for subobjects, and the VTT for the class. They are emitted in a COMDAT group, with the virtual table mangled name as the identifying symbol. Note that if the key function is not declared inline in the class definition, but its definition later is always declared inline, it will be emitted in every object containing the definition.

<b>NOTE</b>: In the abstract, a pure virtual destructor could be used as the key function, as it must be defined even though it is pure. However, the ABI committee did not realize this fact until after the specification of key function was complete; therefore a pure virtual destructor cannot be the key function.

5.2.4 Typeinfo

The RTTI std::type_info structure for a complete class type is emitted in the same object as its virtual table if dynamic, or everywhere referenced if not. The RTTI std::type_info structure for an incomplete class type is emitted wherever referenced. 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 name NTBS objects are emitted with each referencing std::type_info object.

The RTTI std::type_info structures for complete class types and basic types are emitted in COMDAT groups identified by their mangled names. The RTTI std::type_info structures for incomplete class types are emitted with other than the ABI-defined complete type mangled names; an implementation may choose to emit them as local static objects, or in COMDAT groups with implementation-defined names and COMDAT identifiers. The RTTI name NTBS objects are emitted in separate COMDAT groups identified by the NTBS mangled names as weak symbols.

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 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


This version of this document is $Revision: 1.86 $. No special significance should be attached to the form of the revision number; it is simply a identifying number.

[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 bitfields.

[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 trinary 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 (3.3.3).

[000817] Updates from 17 August meeting, email.

[000807] Added base document section (1.5). Further RTTI field name cleanup (2.9.5). Update proposed one-time construction API (3.3.2). Update proposed object construction priority API (3.3.4). 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 (3.3.4).

[000707] Introduce consistent type_info field names (2.9.5). Removed vmi flags for publicly/non-publicly inherited bases (2.9.5). Collect all construction/destruction APIs in one section (3.3). Added one-time initialization API (3.3.2). Vector construction/destruction routines are extern "C" (3.3.3). Added routines for vector construction/destruction (3.3.3). Added copy construction runtime API (3.3.3). 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 bitfield 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 bitfield allocation description (much modified) to the non-virtual-base allocation description. Replace virtual function calling convention description.

[000217] Add excess-size bitfield 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 RTTr 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.