std.experimental.allocator

High-level interface for allocators. Implements bundled allocation/creation and destruction/deallocation of data including structs and classes, and also array primitives related to allocation. This module is the entry point for both making use of allocators and for their documentation.

Synopsis: $(RUNNABLE_EXAMPLE --- // Allocate an int, initialize it with 42 int* p = theAllocator.make!int(42); assert(*p == 42); // Destroy and deallocate it theAllocator.dispose(p); // Allocate using the global process allocator p = processAllocator.make!int(100); assert(*p == 100); // Destroy and deallocate processAllocator.dispose(p); --- ) $(RUNNABLE_EXAMPLE --- // Create an array of 50 doubles initialized to -1.0 double[] arr = theAllocator.makeArray!double(50, -1.0); // Append two zeros to it theAllocator.expandArray(arr, 2, 0.0); // On second thought, take that back theAllocator.shrinkArray(arr, 2); // Destroy and deallocate theAllocator.dispose(arr); --- )

Layered Structure

D's allocators have a layered structure in both implementation and documentation:

1. A high-level, dynamically-typed layer (described further down in this module). It consists of an interface called IAllocator, which concrete allocators need to implement. The interface primitives themselves are oblivious to the type of the objects being allocated; they only deal in void[], by necessity of the interface being dynamic (as opposed to type-parameterized). Each thread has a current allocator it uses by default, which is a thread-local variable theAllocator of type IAllocator. The process has a global allocator called processAllocator, also of type IAllocator. When a new thread is created, processAllocator is copied into theAllocator. An application can change the objects to which these references point. By default, at application startup, processAllocator refers to an object that uses D's garbage collected heap. This layer also include high-level functions such as make and dispose that comfortably allocate/create and respectively destroy/deallocate objects. This layer is all needed for most casual uses of allocation primitives.
2. A mid-level, statically-typed layer for assembling several allocators into one. It uses properties of the type of the objects being created to route allocation requests to possibly specialized allocators. This layer is relatively thin and implemented and documented in the std.experimental.allocator.typed module. It allows an interested user to e.g. use different allocators for arrays versus fixed-sized objects, to the end of better overall performance.
3. A low-level collection of highly generic heap building blocks— Lego-like pieces that can be used to assemble application-specific allocators. The real allocation smarts are occurring at this level. This layer is of interest to advanced applications that want to configure their own allocators. A good illustration of typical uses of these building blocks is module std.experimental.allocator.showcase which defines a collection of frequently- used preassembled allocator objects. The implementation and documentation entry point is std.experimental.allocator.building_blocks. By design, the primitives of the static interface have the same signatures as the IAllocator primitives but are for the most part optional and driven by static introspection. The parameterized class CAllocatorImpl offers an immediate and useful means to package a static low-level allocator into an implementation of IAllocator.
4. Core allocator objects that interface with D's garbage collected heap (std.experimental.allocator.gc_allocator), the C malloc family (std.experimental.allocator.mallocator), and the OS (std.experimental.allocator.mmap_allocator). Most custom allocators would ultimately obtain memory from one of these core allocators.

Idiomatic Use of std.experimental.allocator

As of this time, std.experimental.allocator is not integrated with D's built-in operators that allocate memory, such as new, array literals, or array concatenation operators. That means std.experimental.allocator is opt-in—applications need to make explicit use of it.

For casual creation and disposal of dynamically-allocated objects, use make, dispose, and the array-specific functions makeArray, expandArray, and shrinkArray. These use by default D's garbage collected heap, but open the application to better configuration options. These primitives work either with theAllocator but also with any allocator obtained by combining heap building blocks. For example:

void fun(size_t n)
{
    // Use the current allocator
    int[] a1 = theAllocator.makeArray!int(n);
    scope(exit) theAllocator.dispose(a1);
    ...
}

To experiment with alternative allocators, set theAllocator for the current thread. For example, consider an application that allocates many 8-byte objects. These are not well supported by the default allocator, so a free list allocator would be recommended. To install one in main, the application would use:

void main()
{
    import std.experimental.allocator.building_blocks.free_list
        : FreeList;
    theAllocator = allocatorObject(FreeList!8());
    ...
}

Saving the IAllocator Reference For Later Use

As with any global resource, setting theAllocator and processAllocator should not be done often and casually. In particular, allocating memory with one allocator and deallocating with another causes undefined behavior. Typically, these variables are set during application initialization phase and last through the application.

To avoid this, long-lived objects that need to perform allocations, reallocations, and deallocations relatively often may want to store a reference to the allocator object they use throughout their lifetime. Then, instead of using theAllocator for internal allocation-related tasks, they'd use the internally held reference. For example, consider a user-defined hash table:

struct HashTable
{
    private IAllocator allocator;
    this(size_t buckets, IAllocator allocator = theAllocator) {
        this.allocator = allocator;
        ...
    }
    // Getter and setter
    IAllocator allocator() { return allocator; }
    void allocator(IAllocator a) { assert(empty); allocator = a; }
}

Following initialization, the HashTable object would consistently use its allocator object for acquiring memory. Furthermore, setting HashTable.allocator to point to a different allocator should be legal but only if the object is empty; otherwise, the object wouldn't be able to deallocate its existing state.

Using Allocators without IAllocator

Allocators assembled from the heap building blocks don't need to go through IAllocator to be usable. They have the same primitives as IAllocator and they work with make, makeArray, dispose etc. So it suffice to create allocator objects wherever fit and use them appropriately:

void fun(size_t n)
{
    // Use a stack-installed allocator for up to 64KB
    StackFront!65536 myAllocator;
    int[] a2 = myAllocator.makeArray!int(n);
    scope(exit) myAllocator.dispose(a2);
    ...
}

In this case, myAllocator does not obey the IAllocator interface, but implements its primitives so it can work with makeArray by means of duck typing.

One important thing to note about this setup is that statically-typed assembled allocators are almost always faster than allocators that go through IAllocator. An important rule of thumb is: "assemble allocator first, adapt to IAllocator after". A good allocator implements intricate logic by means of template assembly, and gets wrapped with IAllocator (usually by means of allocatorObject) only once, at client level.