ISO/IEC JTC1 SC22 WG21 N3532 - 2013-03-12
Lawrence Crowl, [email protected], [email protected]
Problem
Solution
Builtin Arrays of Runtime Bound
Presentation
Proposal
Chapter 23 Containers library [containers]
23.2.3 Sequence containers [sequence.reqmts]
23.3 Sequence containers [sequences]
23.3.8 Class template dynarray
[dynarray]
23.3.8.1 Class template dynarray
overview [dynarray.overview]
23.2.8.2 dynarray
constructor and destructor [dynarray.cons]
23.2.8.3 dynarray::size
[dynarray.size]
23.2.8.4 dynarray::data
[dynarray.data]
23.2.8.5 Zero sized dynarrays [dynarray.zero]
Example
Revision History
Programs can become more efficient
when they can bind aspects of their execution
earlier in program development.
As an example,
the std::unordered_map
container
provides more functionality than std::vector
,
but std::vector
provides better performance
when the programmer can bind indexes
to a dense, but extensible, range near zero.
Going further, built-in arrays
provide even better performance
by binding the range end at compilation time.
Unfortunately, for some applications,
the range end is known at container construction
but not at compilation time.
So, built-in arrays are not applicable.
On the other hand,
std::vector
is more general than needed,
as it permits an extensibility that is not required.
Ideally, we would like to be able to specify a container
where the index end is bound at construction,
but does not change thereafter.
The C programming language has such a container in the form of variable-length arrays. They are not general in that they are limited to automatic variables, but given that restriction they are nearly as efficient as normal arrays, requiring only mark/release stack allocation and maintenance of a frame pointer. (Maintaining a frame pointer is a good idea anyway.) Unfortunately the detailed type semantics of C variable-length arrays are probably not acceptable to C++, so we cannot simply adopt them.
The std::valarray
container is intermediate
between built-in arrays and std::vector
,
but as it supports a resize
method,
it cannot hold its size fixed for the lifetime of the variable.
Furthermore, std::valarray
supports compound member assignment operators
that imply such operators in the parameter type.
Such implications are workable only for types with "full interfaces",
not for general types.
Instead of adopting C variable-length arrays,
we propose to define a new facility for arrays
where the number of elements is bound at construction.
We call these dynamic arrays, dynarray
.
In keeping with C++ practice,
we wish to make dynarray
s
usable with more than just automatic variables.
But to take advantage of the efficiency stack allocation,
we wish to make dynarray
optimizable
when used as an automatic variable.
Therefore, we propose to define dynarray
so that compilers can recognize and implement
construction and destruction directly,
without appeal to any particular standard library implementation.
However, to minimize the necessary burden on compilers,
we propose that dynarray
can be implemented as a pure library,
although with lost optimization opportunity.
We believe that the compilers can introduce the optimization without impact on source or binary compatiblity. There may be some change in code profiles and operator new calls as a result of that optimization, but such risks are common to compiler and library upgrades.
Syntactically, our proposal follows the lead of
std::array
and std::vector
containers.
Semantically, our proposal follows the lead of built-in arrays.
That is,
we do not require more out of std::dynarray
element types
than we do of standard array element types.
The dynarray
constructor has a parameter
indicating the number of elements in the container.
Dynarray
requires an element type with a default constructor,
just as the built-in array requires.
Note that dynarray
does not provide a default constructor,
because there is no reasonable default size,
and hence the dynarray
may not take a dynarray
as an element.
Dynarray
provides a copy constructor,
but use of the copy constructor requires that the element type
also have a copy constructor.
The presence of this constructor implies that
users cannot explicitly instantiate the dynarray
template class
on a type that does not have a copy constructor.
This practice already exists in the standard library.
Dynarray
provides random access iterators,
likely implemented as pointers.
The elements must be contiguously allocated,
to enable access via pointer arithmetic.
Dynarray
also provides reverse iterators,
but these definitions imply that
the compiler implementation depends on the standard library implementation,
which is the reverse of the normal dependence.
Dynarray
does not provide
any mechanism for determining whether heap or stack allocation was used.
Dynarray
does not provide
a constructor from an initializer_list
.
An initializer list necessarily has a static number of elements,
and an array or an std::array
is likely more appropriate.
However, there is no technical reason against such a constructor.
Dynarray
does not provide
a constructor from first and last forward iterators.
Such a constructor is possible, though,
as one can determine the size with std::distance(first,last)
.
The technical consideration is that
determining the distance is only constant time for random access iterators.
In
N3497 Runtime-sized arrays with automatic storage duration,
Jens Maurer proposes arrays with runtime bound.
These arrays are to std::dynarray
as normal fixed-size arrays are to std::array
.
There are several similarities and differences.
Both proposals permit allocation on the stack or on the heap, at the discression of the implementation.
Arrays of runtime bound
can be used only for variables of automatic storage duration.
In contrast, dynarray
can be used anywhere.
However, when dynarray
is not used in a automatic variable,
use of the heap is necessarily required.
The types of arrays of runtime bound
are not inspectable (e.g. with decltype
).
In contrast, dynarray
has a normal template class type,
but the bound is not part of that type.
Arrays of runtime bound
throw std::bad_array_length
when the array size is unacceptable.
As currently defined,
dynarray
also throws throw std::bad_array_length
,
under the assumption that
migration between the two is easier if they throw the same exception.
The other choice would be to throw std::length_error
as does std::vector::reserve()
.
Note, however, that the std::vector::vector(size_type n)
does not specify an exception.
The operator new[]
throws std::bad_array_new_length
.
Because it may allocate memory,
dynarray
may also throw std::bad_alloc
.
Arrays of runtime bounds do no allow zero size;
dynarray
does.
One can use initializer lists with arrays of runtime bound.
They are not currently supported with dynarray
.
If dynarray
were to support initializer lists,
it would derive its size from the initializer list.
This approach prevents a specification inconsistency
that is possible in arrays of runtime bound.
Within the proposal, regular code font indicates normative code and variable code font indicates an example implementation. The example implementation is a pure library implementation, and does not include the stack allocation optimization. Thus, the example implementation is a minimal conforming implementation.
Within the example, regular code font indicates example code. There is no use of variable code font.
Within both the proposal and the example, sample font is part of the commentary and not part of either the proposal or the example. This font is usually visually indistinguishable from code font, but should be clear from context.
The code for the definition, implementation and subsequent example
can be extracted from the HTML source
with the following sed
script.
1,/<code class="extract">/ d
/<\/code>/,/<code class="extract">/ d
s|<var>||g
s|</var>||g
s|<|<|g
s|>|>|g
s|&|\&|g
First, to enable a fully compilable implementation and example, we include appropriate library headers and exceptions introduced elsewhere.
#include <stddef.h>
#include <cstring>
#include <limits>
#include <algorithm>
#include <stdexcept>
#include <iostream>
#include <memory>
namespace std { struct bad_array_length { }; }
The dynarray
container definition is as follows.
The section, paragraph, and table references
are based on those of
N3485 Working Draft, Standard for Programming Language C++,
Stefanus Du Toit, November 2012.
Add <dynarray>
to table 87:
Table 87: Containers library summary Subclause Header(s) 23.2 Requirements 23.3 Sequence containers <array>
<deque>
<dynarray>
<forward_list>
<list>
<vector>23.4 Associative containers <map>
<set>23.5 Unordered associative containers <unordered_map>
<unordered_set>23.6 Container adaptors <queue>
<stack>
In table 101, Optional sequence container operations,
add dynarray
to the list of containers
for operations
front
,
back
,
a[n]
, and
at(n)
.
Add a new synopsis:
Header
<dynarray>
synopsisnamespace std { template< class T > struct dynarray; } // namespace std
dynarray
[dynarray]Add a new section.
dynarray
overview [dynarray.overview]Add a new section:
The header
<dynarray>
defines a class template for storing sequences of objects where the size is fixed at construction. Adynarray
supports random access iterators. An instance ofdynarray<T>
stores elements of typeT
. The elements of adynarray
are stored contiguously, meaning that ifd
is andynarray<T>
then it obeys the identity&a[n] == &a[0] + n
for all0 <= n < N
.Unless otherwise specified, all array operations are as described in 23.2. Descriptions are provided here only for operations on
dynarray
that are not described in that clause or for operations where there is additional semantic information.All operations except construction and destruction shall have constant-time complexity.
namespace std { template< class T > struct dynarray { // types: typedef T value_type; typedef T& reference; typedef const T& const_reference; typedef T* iterator; typedef const T* const_iterator; typedef std::reverse_iterator<iterator> reverse_iterator; typedef std::reverse_iterator<const_iterator> const_reverse_iterator; typedef size_t size_type; typedef ptrdiff_t difference_type; // fields: private: T* store; size_type count; // helper functions: void check(size_type n) { if ( n >= count ) throw out_of_range("dynarray"); } T* alloc(size_type n) { if ( n > std::numeric_limits<size_type>::max()/sizeof(T) ) throw std::bad_array_length(); return reinterpret_cast<T*>( new char[ n*sizeof(T) ] ); } public: // construct and destruct: dynarray() = delete; const dynarray operator=(const dynarray&) = delete; explicit dynarray(size_type c) : store( alloc( c ) ), count( c ) { size_type i; try { for ( size_type i = 0; i < count; ++i ) new (store+i) T; } catch ( ... ) { for ( ; i > 0; --i ) (store+(i-1))->~T(); throw; } } dynarray(const dynarray& d) : store( alloc( d.count ) ), count( d.count ) { try { uninitialized_copy( d.begin(), d.end(), begin() ); } catch ( ... ) { delete store; throw; } } ~dynarray() { for ( size_type i = 0; i < count; ++i ) (store+i)->~T(); delete[] store; } // iterators: iterator begin() { return store; } const_iterator begin() const { return store; } const_iterator cbegin() const { return store; } iterator end() { return store + count; } const_iterator end() const { return store + count; } const_iterator cend() const { return store + count; } reverse_iterator rbegin() { return reverse_iterator(end()); } const_reverse_iterator rbegin() const { return reverse_iterator(end()); } reverse_iterator rend() { return reverse_iterator(begin()); } const_reverse_iterator rend() const { return reverse_iterator(begin()); } // capacity: size_type size() const { return count; } size_type max_size() const { return count; } bool empty() const { return count == 0; } // element access: reference operator[](size_type n) { return store[n]; } const_reference operator[](size_type n) const { return store[n]; } reference front() { return store[0]; } const_reference front() const { return store[0]; } reference back() { return store[count-1]; } const_reference back() const { return store[count-1]; } const_reference at(size_type n) const { check(n); return store[n]; } reference at(size_type n) { check(n); return store[n]; } // data access: T* data() { return store; } const T* data() const { return store; } }; } // namespace std
dynarray
constructor and destructor [dynarray.cons]Add a new section:
dynarray(size_type c);
Effects: May or may not invoke the global
operator new
.Throws:
std::bad_array_length
when the size requested is larger than implementable.std::bad_alloc
when there is insufficient memory.
dynarray(const dynarray& d);
Requires: T is Copy Constructible.
Throws:
std::bad_alloc
when there is insufficient memory.Effects: May or may not invoke the global
operator new
.
~dynarray();
Effects: Invokes the global
operator delete
if and only if the constructor invoked the globaloperator new
.
dynarray::size
[dynarray.size]Add a new section:
size_type size() const;
Returns: Returns the argument to the constructor of the object.
size_type max_size() const;
Returns: Returns the argument to the constructor of the object.
bool empty() const;
Returns:
size() == 0
.
dynarray::data
[dynarray.data]Add a new section:
T* data();
T* data() const;
Returns: A pointer to the contiguous storage containing the elements.
Add a new section:
dynarray
shall provide support for the special case of construction with a size of zero. In the case that the size is zero,begin() == end() ==
unique value. The return value ofdata()
is unspecified. The effect of callingfront()
orback()
for a zero-sizeddynarray
is undefined.
Finally, we show a simple set of uses of the container.
Declaring a reference parameter. Using a const iterator.
void dump( const std::dynarray< int > & source )
{
std::dynarray< int >::const_iterator src = source.begin();
for ( ; src != source.end(); src++ )
std::cout << " " << *src;
std::cout << std::endl;
}
Declaring a local dynarray
of computed size.
Using
front
,
back
,
and a non-const iterator.
void lowrap( std::dynarray< int > & target,
const std::dynarray< int > & source )
{
dump( source );
std::dynarray< int > sorted( source );
dump( sorted );
std::sort( sorted.begin(), sorted.end() );
dump( sorted );
const int* srt = &sorted.front();
std::dynarray< int >::iterator tgt( target.begin() );
for ( ; tgt != target.end(); tgt++ ) {
*tgt = *srt;
if ( srt == &sorted.back() )
srt = &sorted.front();
else
srt++;
}
dump( target );
}
Declaring a local dynarray
of fixed size.
Using
size
,
operator[]
,
at
,
index iteration,
and pointer iteration.
int main() {
std::dynarray< int > alpha(8);
std::dynarray< int > gamma(3);
for ( std::dynarray< int >::size_type i = 0; i < gamma.size(); i++ )
gamma[i] = 4 - i;
lowrap( alpha, gamma );
int sum = 0;
for ( std::dynarray< int >::size_type i = 0; i < alpha.size(); i++ )
sum += alpha.at(i);
return sum;
}
This paper revises N2648 = 08-0158 - 2008-05-16 as follows.
Add a section discussing of the relationship to N3497 Runtime-sized arrays with automatic storage duration.
Support zero-length dynarray
s.
Add corresponding specifications
for max_size
and empty
.
Update implementation.
Throw std::bad_array_length
when the size parameter is too large.
Throw std::bad_alloc
when memory allocation fails.
Clarify that there is no support for determing whether heap or stack allocation was used.
Note that a constructor from initializer_list
is not provided but technically feasible.
Note that a constructor from two forward iterators is not provided but technically feasible.
Update the form of the HTML, which affects the script for extracting the code.
Add a 'Revision History' section.