Fixed Size
1. Pass by reference
template <size_t rows, size_t cols>
void process_2d_array_template(int (&array)[rows][cols])
{
std::cout << __func__ << std::endl;
for (size_t i = 0; i < rows; ++i)
{
std::cout << i << ": ";
for (size_t j = 0; j < cols; ++j)
std::cout << array[i][j] << '\t';
std::cout << std::endl;
}
}
In C++ passing the array by reference without losing the dimension information is probably the safest, since one needn't worry about the caller passing an incorrect dimension (compiler flags when mismatching). However, this isn't possible with dynamic (freestore) arrays; it works for automatic (usually stack-living) arrays only i.e. the dimensionality should be known at compile time.
2. Pass by pointer
void process_2d_array_pointer(int (*array)[5][10])
{
std::cout << __func__ << std::endl;
for (size_t i = 0; i < 5; ++i)
{
std::cout << i << ": ";
for (size_t j = 0; j < 10; ++j)
std::cout << (*array)[i][j] << '\t';
std::cout << std::endl;
}
}
The C equivalent of the previous method is passing the array by pointer. This should not be confused with passing by the array's decayed pointer type (3), which is the common, popular method, albeit less safe than this one but more flexible. Like (1), use this method when all the dimensions of the array is fixed and known at compile-time. Note that when calling the function the array's address should be passed process_2d_array_pointer(&a) and not the address of the first element by decay process_2d_array_pointer(a).
Variable Size
These are inherited from C but are less safe, the compiler has no way of checking, guaranteeing that the caller is passing the required dimensions. The function only banks on what the caller passes in as the dimension(s). These are more flexible than the above ones since arrays of different lengths can be passed to them invariably.
It is to be remembered that there's no such thing as passing an array directly to a function in C [while in C++ they can be passed as a reference (1)]; (2) is passing a pointer to the array and not the array itself. Always passing an array as-is becomes a pointer-copy operation which is facilitated by array's nature of decaying into a pointer.
3. Pass by (value) a pointer to the decayed type
// int array[][10] is just fancy notation for the same thing
void process_2d_array(int (*array)[10], size_t rows)
{
std::cout << __func__ << std::endl;
for (size_t i = 0; i < rows; ++i)
{
std::cout << i << ": ";
for (size_t j = 0; j < 10; ++j)
std::cout << array[i][j] << '\t';
std::cout << std::endl;
}
}
Although int array[][10] is allowed, I'd not recommend it over the above syntax since the above syntax makes it clear that the identifier array is a single pointer to an array of 10 integers, while this syntax looks like it's a 2D array but is the same pointer to an array of 10 integers. Here we know the number of elements in a single row (i.e. the column size, 10 here) but the number of rows is unknown and hence to be passed as an argument. In this case there's some safety since the compiler can flag when a pointer to an array with second dimension not equal to 10 is passed. The first dimension is the varying part and can be omitted. See here for the rationale on why only the first dimension is allowed to be omitted.
4. Pass by pointer to a pointer
// int *array[10] is just fancy notation for the same thing
void process_pointer_2_pointer(int **array, size_t rows, size_t cols)
{
std::cout << __func__ << std::endl;
for (size_t i = 0; i < rows; ++i)
{
std::cout << i << ": ";
for (size_t j = 0; j < cols; ++j)
std::cout << array[i][j] << '\t';
std::cout << std::endl;
}
}
Again there's an alternative syntax of int *array[10] which is the same as int **array. In this syntax the [10] is ignored as it decays into a pointer thereby becoming int **array. Perhaps it is just a cue to the caller that the passed array should have at least 10 columns, even then row count is required. In any case the compiler doesn't flag for any length/size violations (it only checks if the type passed is a pointer to pointer), hence requiring both row and column counts as parameter makes sense here.
Note: (4) is the least safest option since it hardly has any type check and the most inconvenient. One cannot legitimately pass a 2D array to this function; C-FAQ condemns the usual workaround of doing int x[5][10]; process_pointer_2_pointer((int**)&x[0][0], 5, 10); as it may potentially lead to undefined behaviour due to array flattening. The right way of passing an array in this method brings us to the inconvenient part i.e. we need an additional (surrogate) array of pointers with each of its element pointing to the respective row of the actual, to-be-passed array; this surrogate is then passed to the function (see below); all this for getting the same job done as the above methods which are more safer, cleaner and perhaps faster.
Here's a driver program to test the above functions:
#include <iostream>
// copy above functions here
int main()
{
int a[5][10] = { { } };
process_2d_array_template(a);
process_2d_array_pointer(&a); // <-- notice the unusual usage of addressof (&) operator on an array
process_2d_array(a, 5);
// works since a's first dimension decays into a pointer thereby becoming int (*)[10]
int *b[5]; // surrogate
for (size_t i = 0; i < 5; ++i)
{
b[i] = a[i];
}
// another popular way to define b: here the 2D arrays dims may be non-const, runtime var
// int **b = new int*[5];
// for (size_t i = 0; i < 5; ++i) b[i] = new int[10];
process_pointer_2_pointer(b, 5, 10);
// process_2d_array(b, 5);
// doesn't work since b's first dimension decays into a pointer thereby becoming int**
}
