[Note: this document is formatted similarly to the SGI STL implementation documentation pages, and refers to concepts and classes defined there. However, neither this document nor the code it describes is associated with SGI, nor is it necessary to have SGI's STL implementation installed in order to use this class.]

sparsetable<T, GROUP_SIZE>

A sparsetable is a Random Access Container that supports constant time random access to elements, and constant time insertion and removal of elements. It implements the "array" or "table" abstract data type. The number of elements in a sparsetable is set at constructor time, though you can change it at any time by calling resize().

sparsetable is distinguished from other array implementations, including the default C implementation, in its stingy use of memory -- in particular, unused array elements require only 1 bit of disk space to store, rather than sizeof(T) bytes -- and by the ability to save and restore contents to disk. On the other hand, this array implementation, while still efficient, is slower than other array implementations.

This class is appropriate for applications that need to store large arrays in memory, or for applications that need these arrays to be persistent.

Example

#include <sparsehash/sparsetable>

using google::sparsetable;      // namespace where class lives by default

sparsetable<int> t(100);
t[5] = 6;
cout << "t[5] = " << t[5];
cout << "Default value = " << t[99];

Definition

Template parameters

Model of

Type requirements

Public base classes

Members

bool operator==(const sparsetable&, const sparsetable&)

bool operator<(const sparsetable&, const sparsetable&)

New members

Notes

[1] sparsetable::destructive_iterator iterates through a sparsetable like a normal iterator, but ++it may delete the element being iterated past. Obviously, this iterator can only be used once on a given table! One application of this iterator is to copy data from a sparsetable to some other data structure without using extra memory to store the data in both places during the copy.

[2] Since operator[] might insert a new element into the sparsetable, it can't possibly be a const member function. In theory, since it might insert a new element, it should cause the element it refers to to become assigned. However, this is undesirable when operator[] is used to examine elements, rather than assign them. Thus, as an implementation trick, operator[] does not really return a reference. Instead it returns an object that behaves almost exactly like a reference. This object, however, delays setting the appropriate sparsetable element to assigned to when it is actually assigned to.

For a bit more detail: the object returned by operator[] is an opaque type which defines operator=, operator reference(), and operator&. The first operator controls assigning to the value. The second controls examining the value. The third controls pointing to the value.

All three operators perform exactly as an object of type reference would perform. The only problems that arise is when this object is accessed in situations where C++ cannot do the conversion by default. By far the most common situation is with variadic functions such as printf. In such situations, you may need to manually cast the object to the right type:

   printf("%d", static_cast<typename table::reference>(table[i]));

Input/Output

It is possible to save and restore sparsetable objects to disk. Storage takes place in two steps. The first writes the table metadata. The second writes the actual data.

To write a sparsetable to disk, first call write_metadata() on an open file pointer. This saves the sparsetable information in a byte-order-independent format.

After the metadata has been written to disk, you must write the actual data stored in the sparsetable to disk. If the value is "simple" enough, you can do this by calling write_nopointer_data(). "Simple" data is data that can be safely copied to disk via fwrite(). Native C data types fall into this category, as do structs of native C data types. Pointers and STL objects do not.

Note that write_nopointer_data() does not do any endian conversion. Thus, it is only appropriate when you intend to read the data on the same endian architecture as you write the data.

If you cannot use write_nopointer_data() for any reason, you can write the data yourself by iterating over the sparsetable with a const_nonempty_iterator and writing the key and data in any manner you wish.

To read the hashtable information from disk, first you must create a sparsetable object. Then open a file pointer to point to the saved sparsetable, and call read_metadata(). If you saved the data via write_nopointer_data(), you can follow the read_metadata() call with a call to read_nopointer_data(). This is all that is needed.

If you saved the data through a custom write routine, you must call a custom read routine to read in the data. To do this, iterate over the sparsetable with a nonempty_iterator; this operation is sensical because the metadata has already been set up. For each iterator item, you can read the key and value from disk, and set it appropriately. The code might look like this:

   for (sparsetable<int*>::nonempty_iterator it = t.nonempty_begin();
        it != t.nonempty_end(); ++it) {
       *it = new int;
       fread(*it, sizeof(int), 1, fp);
   }

Here's another example, where the item stored in the sparsetable is a C++ object with a non-trivial constructor. In this case, you must use "placement new" to construct the object at the correct memory location.

   for (sparsetable<ComplicatedCppClass>::nonempty_iterator it = t.nonempty_begin();
        it != t.nonempty_end(); ++it) {
       int constructor_arg;   // ComplicatedCppClass takes an int to construct
       fread(&constructor_arg, sizeof(int), 1, fp);
       new (&(*it)) ComplicatedCppClass(constructor_arg);     // placement new
   }

See also

The following are SGI STL concepts and classes related to sparsetable.