In lesson 4.5 -- Unsigned integers, and why to avoid them, we noted how we generally prefer to use signed values to hold quantities, because unsigned values can act in surprising ways. However, in lesson 16.3 -- std::vector and the unsigned length and subscript problem, we discussed how std::vector
(and other container classes) uses unsigned integral type std::size_t
for length and indices.
This can lead to problems such as this one:
#include <iostream>
#include <vector>
template <typename T>
void printReverse(const std::vector<T>& arr)
{
for (std::size_t index{ arr.size() - 1 }; index >= 0; --index) // index is unsigned
{
std::cout << arr[index] << ' ';
}
std::cout << '\n';
}
int main()
{
std::vector arr{ 4, 6, 7, 3, 8, 2, 1, 9 };
printReverse(arr);
return 0;
}
This code begins by printing the array in reverse:
9 1 2 8 3 7 6 4
And then exhibits undefined behavior. It might print garbage values, or crash the application.
There are two problems here. First, our loop executes as long as index >= 0
(or in other words, as long as index
is positive), which is always true when index
is unsigned. Therefore, the loop never terminates.
Second, when we decrement index
when it has value 0
, it will wrap around to a large positive value, which we then use to index the array on the next iteration. This is an out-of-bounds index, and will cause undefined behavior. We run into the same problem if our vector is empty.
And while there are plenty of ways to work around these specific issues, these kinds of issues are magnets for bugs.
Using a signed type for a loop variable more easily avoids such problems, but has its own challenges. Here’s a version of the above problem that uses a signed index:
#include <iostream>
#include <vector>
template <typename T>
void printReverse(const std::vector<T>& arr)
{
for (int index{ static_cast<int>(arr.size()) - 1}; index >= 0; --index) // index is signed
{
std::cout << arr[static_cast<std::size_t>(index)] << ' ';
}
std::cout << '\n';
}
int main()
{
std::vector arr{ 4, 6, 7, 3, 8, 2, 1, 9 };
printReverse(arr);
return 0;
}
While this version functions as intended, the code is also a cluttered due to the addition two static casts. arr[static_cast<std::size_t>(index)]
is particularly hard to read. In this case, we’ve improved safety at a significant cost to readability.
Here’s another example of using a signed index:
#include <iostream>
#include <vector>
// Function template to calculate the average value in a std::vector
template <typename T>
T calculateAverage(const std::vector<T>& arr)
{
int length{ static_cast<int>(arr.size()) };
T average{ 0 };
for (int index{ 0 }; index < length; ++index)
average += arr[static_cast<std::size_t>(index)];
average /= length;
return average;
}
int main()
{
std::vector testScore1 { 84, 92, 76, 81, 56 };
std::cout << "The class 1 average is: " << calculateAverage(testScore1) << '\n';
return 0;
}
The cluttering of our code with static casts is pretty terrible.
So what should we do? This is an area where there is no ideal solution.
There are many viable options here, which we’ll present in order from what we believe is worst to best. You will likely encounter all of these in code written by others.
Author’s note
Although we’ll be discussing this in the context of std::vector
, all of the standard library containers (e.g. std::array
) work similarly and have the same challenges. The discussion that follows is applicable to any of them.
Leave signed/unsigned conversion warnings off
If you were wondering why signed/unsigned conversion warnings are often disabled by default, this topic is one of the key reasons. Every time we subscript a standard library container using a signed index, a sign conversion warning will be generated. This will quickly fill up your compilation log with spurious warnings, drowning out warnings that may actually be legitimate.
So one way to avoid having to deal with lots of signed/unsigned conversion warnings is to simply leave those warnings turned off.
This is the simplest solution, but not one we recommend, as this will also suppress generation of legitimate sign conversion warnings that may cause bugs if not addressed.
Using an unsigned loop variable
Many developers believe that since the standard library array types were designed to use unsigned indices, then we should use unsigned indices! This is a completely reasonable position. We just need to be extra careful that we do not run into signed/unsigned mismatches when doing so. If possible, avoid using the index loop variable for anything but indexing.
If we decide to use this approach, which unsigned type should we actually use?
In lesson 16.3 -- std::vector and the unsigned length and subscript problem, we noted that the standard library container classes define nested typedef size_type
, which is an unsigned integral type used for array lengths and indices. The size()
member function returns size_type
, and operator[]
uses size_type
as an index, so using size_type
as the type of your index is technically the most consistent and safe unsigned type to use (as it will work in all cases, even in the extremely rare case where size_type
is something other than size_t
.). For example:
#include <iostream>
#include <vector>
int main()
{
std::vector arr { 1, 2, 3, 4, 5 };
for (std::vector<int>::size_type index { 0 }; index < arr.size(); ++index)
std::cout << arr[index] << ' ';
return 0;
}
However, using size_type
has a major downside: because it is a nested type, to use it we have to explicitly prefix the name with the fully templated name of the container (meaning we have to type std::vector<int>::size_type
rather than just std::size_type
). This requires a lot of typing, is hard to read, and varies depending on the container and element type.
When used inside a function template, we can use T
for the template arguments. But we also need to prefix the type with the typename
keyword:
#include <iostream>
#include <vector>
template <typename T>
void printArray(const std::vector<T>& arr)
{
// typename keyword prefix required for dependent type
for (typename std::vector<T>::size_type index { 0 }; index < arr.size(); ++index)
std::cout << arr[index] << ' ';
}
int main()
{
std::vector arr { 9, 7, 5, 3, 1 };
printArray(arr);
return 0;
}
If you forget the typename
keyword, your compiler will probably remind you to add it.
For advanced readers
Any name that depends on a type containing a template parameter is called a dependent name. Dependent names must be prefixed with the keyword typename
in order to be used as a type.
In the above example, std::vector<T>
is a type with a template parameter, so nested type std::vector<T>::size_type
is a dependent name, and must be prefixed with typename
to be used as a type.
You may occasionally see the array type aliased to make the loop easier to read:
using arrayi = std::vector<int>;
for (arrayi::size_type index { 0 }; index < arr.size(); ++index)
A more general solution is to have the compiler fetch the type of the array type object for us, so that we don’t have to explicitly specify the container type or template arguments. To do so, we can use the decltype keyword, which returns the type of its parameter.
// arr is some non-reference type
for (decltype(arr)::size_type index { 0 }; index < arr.size(); ++index) // decltype(arr) resolves to std::vector<int>
However, if arr
is a reference type (e.g. an array passed by reference), the above doesn’t work. We need to first remove the reference from arr
:
template <typename T>
void printArray(const std::vector<T>& arr)
{
// arr can be a reference or non-reference type
for (typename std::remove_reference_t<decltype(arr)>::size_type index { 0 }; index < arr.size(); ++index)
std::cout << arr[index] << ' ';
}
Unfortunately, this is no longer very concise or easy to remember.
Because size_type
is almost always a typedef for size_t
, many programmers just skip using size_type
altogether and use the easier to remember and type std::size_t
directly:
for (std::size_t index { 0 }; index < arr.size(); ++index)
Unless you’re using custom allocators (and you probably aren’t), we believe this is a reasonable approach.
Using a signed loop variable
Although it makes working with the standard library container types a bit more difficult, using a signed loop variable is consistent with the best practices employed in the rest of our code (to favor signed values for quantities). And the more we can consistently apply our best practices, the fewer errors we will have overall.
If we are going to use signed loop variables, there are three issues we need to address:
- What signed type should we use?
- Getting the length of the array as a signed value
- Converting the signed loop variable to an unsigned index
What signed type should we use?
There are three (sometimes four) good options here.
- Unless you are working with a very large array, using
int
should be fine (particularly on architectures where int is 4 bytes).int
is the default signed integral type we use for everything when we don’t really care about the type otherwise, and there’s little reason to do otherwise here. - If you are dealing with very large arrays, or if you want to be a bit more defensive, you can use the strangely named
std::ptrdiff_t
. This typedef is often used as the signed counterpart tostd::size_t
. - Because
std::ptrdiff_t
has a weird name, another good approach is to define your own type alias for indices:
using Index = std::ptrdiff_t;
// Sample loop using index
for (Index index{ 0 }; index < static_cast<Index>(arr.size()); ++index)
We’ll show a full example of this in the next section.
Defining your own type alias also has a potential future benefit: if the C++ standard library ever releases a type designed to be used as a signed index, it will be easy to either modify Index
to alias that type, or to find/replace Index
with whatever that type is named.
- In cases where you can derive the type of your loop variable from the initializer, you can use
auto
to have the compiler deduce the type:
for (auto index{ static_cast<std::ptrdiff_t>(arr.size())-1 }; index >= 0; --index)
In C++23, the Z
suffix can be used to define a literal of the type that is the signed counterpart to std::size_t
(probably std::ptrdiff_t
):
for (auto index{ 0Z }; index < static_cast<std::ptrdiff_t>(arr.size()); ++index)
Getting the length of an array as a signed value
- Pre-C++20, the best option is to
static_cast
the return value of thesize()
member function orstd::size()
to a signed type:
#include <iostream>
#include <vector>
using Index = std::ptrdiff_t;
int main()
{
std::vector arr{ 9, 7, 5, 3, 1 };
for (auto index{ static_cast<Index>(arr.size())-1 }; index >= 0; --index)
std::cout << arr[static_cast<std::size_t>(index)] << ' ';
return 0;
}
That way, the unsigned value returned by arr.size()
will be converted to a signed type, so our comparison operator will have two signed operands. And because signed indices won’t overflow when they go negative, we don’t have the wrap-around problem we ran into when using unsigned indices.
The downside of this approach is that it clutters up our loop, making it harder to read. We can address this by moving the length out of the loop:
#include <iostream>
#include <vector>
using Index = std::ptrdiff_t;
int main()
{
std::vector arr{ 9, 7, 5, 3, 1 };
auto length{ static_cast<Index>(arr.size()) };
for (auto index{ length-1 }; index >= 0; --index)
std::cout << arr[static_cast<std::size_t>(index)] << ' ';
return 0;
}
- In C++20, use
std::ssize()
:
If you want more evidence that the designers of C++ now believe that signed indices are the way to go, consider the introduction of std::ssize()
in C++20. This function returns the size of an array type as a signed type (likely ptrdiff_t
).
#include <iostream>
#include <vector>
int main()
{
std::vector arr{ 9, 7, 5, 3, 1 };
for (auto index{ std::ssize(arr)-1 }; index >= 0; --index) // std::ssize introduced in C++20
std::cout << arr[static_cast<std::size_t>(index)] << ' ';
return 0;
}
Converting the signed loop variable to an unsigned index
Once we have a signed loop variable, we’re going to run into implicit sign conversion warnings whenever we try to use that signed loop variable as an index. So we need some way to convert our signed loop variable to an unsigned value wherever we intend to use it as an index.
- The obvious option is to static cast our signed loop variable into an unsigned index. We show this in the prior example. Unfortunately, we need to do this everywhere we subscript the array, and it makes our array indices hard to read.
- Use a conversion function with a short name:
#include <iostream>
#include <type_traits> // for std::is_integral and std::is_enum
#include <vector>
using Index = std::ptrdiff_t;
// Helper function to convert `value` into an object of type std::size_t
// UZ is the suffix for literals of type std::size_t.
template <typename T>
constexpr std::size_t toUZ(T value)
{
// make sure T is an integral type
static_assert(std::is_integral<T>() || std::is_enum<T>());
return static_cast<std::size_t>(value);
}
int main()
{
std::vector arr{ 9, 7, 5, 3, 1 };
auto length { static_cast<Index>(arr.size()) }; // in C++20, prefer std::ssize()
for (auto index{ length-1 }; index >= 0; --index)
std::cout << arr[toUZ(index)] << ' '; // use toUZ() to avoid sign conversion warning
return 0;
}
In the above example, we’ve created a function named toUZ()
that is designed to convert integral values to values of type std::size_t
. This allows us to index our array as arr[toUZ(index)]
, which is pretty readable.
- Use a custom view
In prior lessons, we discussed how std::string
owns a string, whereas std::string_view
is a view into a string that exists elsewhere. One of the neat things about std::string_view
is how it can view different types of strings (C-style string literals, std::string
, and other std::string_view
) but keeps a consistent interface for us to use.
While we can’t modify the standard library containers to accept a signed integral index, we can create our own custom view class to “view” a standard library container class. And in doing so, we can define our own interface to work however we want.
In the following example, we define a custom view class that can view any standard library container that supports indexing. Our interface will do two things:
- Allow us to access elements using
operator[]
with a signed integral type. - Get the length of the container as a signed integral type (since
std::ssize()
is only available on C++20).
This uses operator overloading (which we briefly introduced in lesson 13.5 -- Introduction to overloading the I/O operators) to implement operator[]
. You don’t need to know how SignedArrayView
is implemented in order to use it.
SignedArrayView.h:
#ifndef SIGNED_ARRAY_VIEW_H
#define SIGNED_ARRAY_VIEW_H
#include <cstddef> // for std::size_t and std::ptrdiff_t
// SignedArrayView provides a view into a container that supports indexing
// allowing us to work with these types using signed indices
template <typename T>
class SignedArrayView // requires C++17
{
private:
T& m_array;
public:
using Index = std::ptrdiff_t;
SignedArrayView(T& array)
: m_array{ array } {}
// Overload operator[] to take a signed index
constexpr auto& operator[](Index index) { return m_array[static_cast<typename T::size_type>(index)]; }
constexpr const auto& operator[](Index index) const { return m_array[static_cast<typename T::size_type>(index)]; }
constexpr auto ssize() const { return static_cast<Index>(m_array.size()); }
};
#endif
main.cpp:
#include <iostream>
#include <vector>
#include "SignedArrayView.h"
int main()
{
std::vector arr{ 9, 7, 5, 3, 1 };
SignedArrayView sarr{ arr }; // Create a signed view of our std::vector
for (auto index{ sarr.ssize() - 1 }; index >= 0; --index)
std::cout << sarr[index] << ' '; // index using a signed type
return 0;
}
Index the underlying C-style array instead
In lesson 16.3 -- std::vector and the unsigned length and subscript problem, we noted that instead of indexing the standard library container, we can instead call the data()
member function and index that instead. Since data()
returns the array data as a C-style array, and C-style arrays allow indexing with both signed and unsigned values, this avoids sign conversion issues.
int main()
{
std::vector arr{ 9, 7, 5, 3, 1 };
auto length { static_cast<Index>(arr.size()) }; // in C++20, prefer std::ssize()
for (auto index{ length - 1 }; index >= 0; --index)
std::cout << arr.data()[index] << ' '; // use data() to avoid sign conversion warning
return 0;
}
We believe that this method is the best of the indexing options:
- We can use signed loop variables and indices.
- We don’t have to define any custom types or type aliases.
- The hit to readability from using
data()
isn’t very big. - There should be no performance hit in optimized code.
The only sane choice: avoid indexing altogether!
All of the options presented above have their own downsides, so it’s hard to recommend one approach over the other. However, there is a choice that is far more sane than the others: avoid indexing with integral values altogether.
C++ provides several other methods for traversing through arrays that do not use indices at all. And if we don’t have indices, then we don’t run into all of these signed/unsigned conversion issues.
Two common methods for array traversal without indices include range-based for loops, and iterators.
Related content
We cover ranged-for loops in the next lesson (16.8 -- Range-based for loops (for-each)).
We cover iterators in upcoming lesson 18.2 -- Introduction to iterators.
If you’re only using the index variable to traverse the array, then prefer a method that does not use indices.
Best practice
Avoid array indexing with integral values whenever possible.