Runtime-sized arrays with automatic storage duration (revision 5)

Status

Changes compared to N3366

• discuss array of runtime bound in range-based for statement
• multidimensional arrays with multiple runtime bounds is an open issue
• adjust grammar in 8 dcl.decl
• for "bound is 0", range-checking of bound, and "too many initializers", behave as-if in array-new

Changes compared to N3412

• removed open issues after EWG discussion: no support for range-based for, only have the outermost dimension as a non-constant
• rename exception from std::bad_array_new_length to std::bad_array_length
• permitted allocation on heap
• change 4.2 conv.array to accomodate arrays with no elements

Changes compared to N3467

• removed support for zero-sized arrays (CWG consensus to defer that to a separate proposal that also addresses the questions of zero object size)
• prohibit taking the address of a runtime-sized array

Changes compared to N3497

• cv-qualification adjustment is also applied for runtime-sized arrays
• range-based "for" loops are supported, also inside lambdas

Motivation

Sometimes, a user wishes to allocate a local array whose size is not known at compile-time, but at runtime only. Nonetheless, the array's size will remain unchanged during the lifetime of the array.

Examples are

• interfacing with the readv/writev POSIX system calls
• small runtime-sized data structure to apply STL algorithms calls
• ...

This paper proposes to add local runtime-sized arrays with automatic storage duration to C++, for example:

void f(std::size_t n)
{
   int a[n];
   for (std::size_t i = 0; i < n; ++i)
       a[i] = 2*i;
   std::sort(a, a+n);
}

Traditionally, the array bound "n" had to be a constant expression (see 8.3.4 dcl.array). For local arrays with automatic storage duration, this paper proposes to lift that restriction. The syntax is intended to be the same as that used for C99 variable length arrays (VLAs), except that implementation support for VLAs is optional in C11 and aggregate initialization is not supported in C.

As a design guideline, the same rules should apply to "new T[n]" and a local array "T a[n]", except that arrays of zero size are only supported for new.

There is well-established existing practice with gcc, Clang, and Intel C++ all implementing a similar, if not identical feature. In fact, Douglas Gregor reported in c++std-ext-12553 on 2012-01-30:

Users really seem to want this feature. It's a fairly common extension, and when we tried to ban it out of principle (in Clang), our users reacted *very* strongly.

Scope

• multidimensional arrays, where other than the top level has a runtime bound (in analogy, array-new doesn't support that either)
• modifications to the function declarator syntax
• sizeof(a) being a runtime-evaluated expression returning the size of a
• "typedef int a[n];" evaluating "n" and passing that through the typedef

void f(std::size_t n)
{
  int a[n];
  unsigned int x = sizeof(a);            // ill-formed
  const std::type_info& ti = typeid(a);  // ill-formed
  typedef int t[n];                      // ill-formed
}

Discussion

Data structures that allocate from the heap access, by design, a global resource that is often highly contended in a multi-threaded program. Therefore, avoiding heap allocations is usually advantageous for performance. Allocating such data from the stack is much more efficient, because the stack is local to each thread and bytes on the stack are often cached locally. (Since each thread has a separate stack, it is unlikely that another thread on another CPU accesses the same data, thereby causing more expensive cache invalidations.)

The syntax does not require additional keywords. Instead, a restriction on the existing array declaration syntax is lifted in certain circumstances.

There is no reason to limit the feature to PODs as array element types, thus such a limitation is not proposed.

Stack overflow becomes more likely, in particular if the size depends on external input and is not properly checked. Some environments might therefore prohibit the use of the feature. Such a prohibition can be easily enforced with a static analysis tool.

There is no longer an upper bound on the size of a function's stack frame. This makes static analysis of stack usage harder.

Alternatives

• std::dynarray<T> as presented in N2648 "C++ Dynamic Arrays" by Lawrence Crowl and Matt Austern: stack allocation is left as QoI, requiring "compiler detection" when using dynarray<T> locally; also usable with static storage duration
• alloca(): not type-safe, no support for C++ constructors/destructors
• std::vector: allocates from heap, allows space "reservation", thus requiring more management data
• std::valarray: allocates from heap; additional unrelated requirements on contained type (see 26.2 numeric.requirements)
• "auto_buffer" (fixed-size local buffer plus overflow to the heap): may waste memory for fixed-size buffer; introduces additional layer of pointer indirection

Changes to the Working Draft

A type is a literal type if it is:

• ...
• an array of literal type other than an array of runtime bound; or
• ...

These methods of constructing types can be applied recursively; restrictions are mentioned in 8.3.1 dcl.ptr, 8.3.4 dcl.array, 8.3.5 dcl.fct, and 8.3.2 dcl.ref. Constructing a type such that the number of bytes in its object representation exceeds the maximum value representable in the type std::size_t (18.2 support.types) is ill-formed.

An lvalue or rvalue expression of type "array of N T", "array of runtime bound of T", or "array of unknown bound of T" can be converted to a prvalue of type "pointer to T". The result is a pointer to the first element of the array.

[ Note: ... ] An array of runtime bound (8.3.4 dcl.array) shall not be captured by copy.

An entity is captured by reference if it is implicitly or explicitly captured but not captured by copy. It is unspecified whether additional unnamed non-static data members are declared in the closure type for entities captured by reference. [ Note: Capturing by reference an array of runtime bound also implicitly captures the value of the bound to support the range-based for statement (6.5.4 stmt.ranged).]

The typeid operator shall not be applied to an array of runtime bound.

When typeid is applied to a glvalue expression ...

The result of the unary & operator is a pointer to its operand. The operand shall be an lvalue of type other than "array of runtime bound" or a qualified-id. ...

... The sizeof operator shall not be applied to an expression that has function or incomplete type, to an enumeration type whose underlying type is not fixed before all its enumerators have been declared, to an array of runtime bound, to the parenthesized name of such types, or to an lvalue that designates a bit-field. ...

Drafting note: 5.3.7 expr.unary.noexcept does not need to be changed, because the declaration of an array of runtime bound cannot be lexically part of the operand of a noexcept; see also 5.1.2p2 expr.prim.lambda.

Insert a new paragraph before 7.1.3 dcl.typedef paragraph 3:

A typedef-name shall not name an array of runtime bound.

In a given non-class scope, a typedef specifier can be used to redefine the name of any type declared in that scope to refer to the type to which it already refers. [ Example: ... ]

The type denoted by decltype(e) is defined as follows:

• if e has type "array of runtime bound", the program is ill-formed;
• otherwise, if e is an unparenthesized ...

noptr-declarator:
      declarator-id attribute-specifier-seqopt
      noptr-declarator parameters-and-qualifiers
      noptr-declarator [ constant-expressionopt expressionopt ] attribute-specifier-seqopt
      ( ptr-declarator )

Drafting note: Section 8.1 [dcl.name] defining the grammar term type-id is intentionally unchanged. Thus, constructing an array of runtime bound in a type-id is ill-formed, because the grammar continues to require all constant-expressions in array bounds.

... Similarly, the optional attribute-specifier-seq (7.6.1) appertains to the pointer and not to the object pointed to. There shall be no pointers to arrays of runtime bound.

There shall be no references to references, no references to arrays of runtime bound, no arrays of references, and no pointers to references. ...

In a declaration T D where D has the form

           D1 [ constant-expressionopt expressionopt ] attribute-specifier-seqopt

and the type of the identifier in the declaration T D1 is "derived-declarator-type-list T", then the type of the identifier of D is an array type; if the type of the identifier of D contains the auto type-specifier, the program is ill-formed. T is called the array element type; this type shall not be a reference type, the (possibly cv-qualified) type void, a function type, an array of unknown or runtime bound, or an abstract class type. Except as noted below, if the expression is omitted, the type of the identifier of D is "derived-declarator-type-list array of unknown bound of T". If the expression is present, it is implicitly converted to std::size_t. The expression is erroneous if:

• its value before converting to std::size_t or, in the case of an expression of class type, before application of the second standard conversion (13.3.3.1.2 over.ics.user) is less than or equal to zero;
• its value is such that the size of the allocated object would exceed the implementation-defined limit (annex B implimits);
• the initializer of the object is a braced-init-list whose number of (top-level) initializer-clauses exceeds the number of elements to initialize; or
• the object is initialized with a string literal and there are more initializers than there are array elements.

If the expression, after converting to std::size_t, is a core constant expression and the expression is erroneous, the program is ill-formed.

• If the expression, after converting to std::size_t, is a core constant expression whose value is N, the type of the identifier of D is "derived-declarator-type-list array of N T".
• Otherwise, the type of the identifier of D is "derived-declarator-type-list array of runtime bound of T" and the value of the expression designates the number of elements N in the array. If the expression is erroneous, an exception of a type that would match a handler (15.3 except.handle) of type std::bad_array_length (18.6.2.2 xxx) is thrown.

If the constant-expression (5.19 expr.const) is present, it shall be an integral constant expression and its value shall be greater than zero. The constant expression specifies the bound of (number of elements in) the array. If the value of the constant expression is N, the array has N elements numbered 0 to N-1, and the type of the identifier of D is "derived-declarator-type-list array of N T". An object of array type contains a contiguously allocated non-empty set of N subobjects of type T. Except as noted below, if the constant expression is omitted, the type of the identifier of D is "derived-declarator-type-list array of unknown bound of T", an incomplete object type. The type "derived-declarator-type-list array of N T" is a different type from the type "derived-declarator-type-list array of unknown bound of T", see 3.9 basic.types. Any type of the form "cv-qualifier-seq array of N T" is adjusted to "array of N cv-qualifier-seq T", and similarly for "array of unknown bound of T" and "array of runtime bound of T". The optional attribute-specifier-seq appertains to the array. [ Example:

  typedef int A[5], AA[2][3];
  typedef const A CA;         // type is "array of 5 const int"
  typedef const AA CAA;       // type is "array of 2 array of 3 const int"

  void f(unsigned int n) {
    int a[n];     // type of "a" is "array of runtime bound of int"
  }

-- end example ] [ Note: ... ]

When several "array of" specifications are adjacent, a multidimensional array is created; only the first of the constant expressions that specify the bounds of the arrays may be omitted. In addition to ...

An array of runtime bound shall only be used as the type of a local object with automatic storage duration. If the size of the array exceeds the size of the memory available for objects with automatic storage duration, the behavior is undefined [ Footnote: Implementations that detect this case are encouraged to throw an exception that would match a handler (15.3 except.handle) of type std::bad_array_length (18.6.2.2 xxx). ] It is unspecified whether a global allocation function (3.7.4 basic.stc.dynamic) is invoked to obtain storage for the array. If it is invoked, the corresponding global deallocation function is invoked to release the storage after the lifetime of the array ended. [ Footnote: Alternatively, an implementation could allocate such an array on the usual stack or obtain storage via malloc (20.6.13 c.malloc). ]

If the type of a parameter includes a type of the form "array of runtime bound of T", "pointer to array of unknown bound of T", or "reference to array of unknown bound of T," the program is ill-formed. [ Footnote: ... ] Functions shall not have a return type of type array or function, although they may have a return type of type pointer or reference to such things.

For types other than arrays of runtime bound (8.3.4 dcl.array), an An initializer-list is ill-formed if the number of initializer-clauses exceeds the number of members or elements to initialize. [ Example: ... ]

[ Note: There cannot be more initializers than there are array elements; see 8.3.4 dcl.array. [ Example:

 char cv[4] = "asdf";                 // error

is ill-formed since there is no space for the implied trailing '\0'. -- end example ] -- end note ]

Non-static A non-static (9.4 class.static) data members member shall not have incomplete types. type or type "array of runtime bound". [ Note: In particular, a class C shall not contain a non-static member of class C, but it can contain a pointer or reference to an object of class C. ]

A non-type template-parameter shall not be declared to have floating point, class, array of runtime bound, or void type. [ Example: ... ]

Drafting note: It is not necessary to explicitly prevent template argument deduction for an array of runtime bound, because 14.8.2p8 [temp.deduct] says "If a substitution results in an invalid type or expression, type deduction fails. An invalid type or expression is one that would be ill-formed if written using the substituted arguments." 8.3.5p8 (and other places) establishing restrictions on forming types are thus sufficient.

... [ Note: An exception can be thrown from one of the following contexts: throw-expression (see below), allocation functions (3.7.4.1 basic.stc.dynamic.allocation), dynamic_cast (5.2.7 expr.dynamic.cast), typeid (5.2.8 expr.typeid), new-expression (5.3.4 expr.new), array of runtime bound (8.3.4 dcl.array), and standard library functions (17.5.1.4 structure.specifications). -- end note ] ...

Class bad_array_length

  namespace std {
     class bad_array_length : public bad_alloc {
     public:
        bad_array_length() noexcept;
     };
  }

The class bad_array_length defines the type of objects thrown as exceptions by the implementation to report an attempt to allocate an array of runtime bound with a size less than or equal to zero or greater than an implementation-defined limit (8.3.4 dcl.array).

bad_array_length() noexcept;

Effects: constructs an object of class bad_array_length.
Remarks: the result of calling what() on the newly constructed object is implementation-defined.