Constructs a block allocator given a hunk of memory, or a desired capacity in bytes.
If ParentAllocator is not NullAllocator and defines deallocate, the destructor is defined to deallocate the block held.
The alignment offered is user-configurable statically through parameter theAlignment, defaulted to platformAlignment.
If blockSize == chooseAtRuntime, BitmappedBlock offers a read/write property blockSize. It must be set before any use of the allocator. Otherwise (i.e. theBlockSize is a legit constant), blockSize is an alias for theBlockSize. Whether constant or variable, must also be a multiple of alignment. This constraint is asserted statically and dynamically.
Allocates a block with specified alignment a. The alignment must be a power of 2. If a <= alignment, function forwards to allocate. Otherwise, it attempts to overallocate and then adjust the result for proper alignment. In the worst case the slack memory is around two blocks.
Reallocates a block previously allocated with alignedAllocate. Contractions do not occur in place.
Allocates s bytes of memory and returns it, or null if memory could not be allocated.
If the BitmappedBlock object is empty (has no active allocation), allocates all memory within and returns a slice to it. Otherwise, returns null (i.e. no attempt is made to allocate the largest available block).
Allocates s bytes of memory and returns it, or null if memory could not be allocated. allocateFresh behaves just like allocate, the only difference being that this always returns unused(fresh) memory. Although there may still be available space in the BitmappedBlock, allocateFresh could still return null, because all the available blocks have been previously deallocated.
Deallocates a block previously allocated with this allocator.
Forcibly deallocates all memory allocated by this allocator, making it available for further allocations. Does not return memory to ParentAllocator.
Returns Ternary.yes if no memory is currently allocated with this allocator, otherwise Ternary.no. This method never returns Ternary.unknown.
Expands in place a buffer previously allocated by BitmappedBlock. If instantiated with No.multiblock, the expansion fails if the new length exceeds theBlockSize.
Returns the actual bytes allocated when n bytes are requested, i.e. n.roundUpToMultipleOf(blockSize).
Returns Ternary.yes if b belongs to the BitmappedBlock object, Ternary.no otherwise. Never returns Ternary.unkown. (This method is somewhat tolerant in that accepts an interior slice.)
Reallocates a previously-allocated block. Contractions occur in place.
The parent allocator. Depending on whether ParentAllocator holds state or not, this is a member variable or an alias for ParentAllocator.instance.
the length of a block, which must be a multiple of theAlignment
alignment of each block
allocator from which the BitmappedBlock will draw memory. If set to NullAllocator, the storage must be passed via the constructor
Yes.multiblock to support allocations spanning across multiple blocks and No.multiblock to support single block allocations. Although limited by single block allocations, No.multiblock will generally provide higher performance.
// Create a block allocator on top of a 10KB stack region. import std.experimental.allocator.building_blocks.region : InSituRegion; import std.traits : hasMember; InSituRegion!(10_240, 64) r; auto a = BitmappedBlock!(64, 64)(cast(ubyte[])(r.allocateAll())); static assert(hasMember!(InSituRegion!(10_240, 64), "allocateAll")); const b = a.allocate(100); assert(b.length == 100);
import std.experimental.allocator.mallocator : Mallocator; import std.typecons : Flag, Yes; enum blockSize = 64; enum numBlocks = 10; // The 'BitmappedBlock' is implicitly instantiated with Yes.multiblock auto a = BitmappedBlock!(blockSize, 8, Mallocator, Yes.multiblock)(numBlocks * blockSize); // Instantiated with Yes.multiblock, can allocate more than one block at a time void[] buf = a.allocate(2 * blockSize); assert(buf.length == 2 * blockSize); assert(a.deallocate(buf)); // Can also allocate less than one block buf = a.allocate(blockSize / 2); assert(buf.length == blockSize / 2); // Expands inside the same block assert(a.expand(buf, blockSize / 2)); assert(buf.length == blockSize); // If Yes.multiblock, can expand past the size of a single block assert(a.expand(buf, 3 * blockSize)); assert(buf.length == 4 * blockSize); assert(a.deallocate(buf));
import std.experimental.allocator.mallocator : Mallocator; import std.typecons : Flag, No; enum blockSize = 64; auto a = BitmappedBlock!(blockSize, 8, Mallocator, No.multiblock)(1024 * blockSize); // Since instantiated with No.multiblock, can only allocate at most the block size void[] buf = a.allocate(blockSize + 1); assert(buf is null); buf = a.allocate(blockSize); assert(buf.length == blockSize); assert(a.deallocate(buf)); // This is also fine, because it's less than the block size buf = a.allocate(blockSize / 2); assert(buf.length == blockSize / 2); // Can expand the buffer until its length is at most 64 assert(a.expand(buf, blockSize / 2)); assert(buf.length == blockSize); // Cannot expand anymore assert(!a.expand(buf, 1)); assert(a.deallocate(buf));
BitmappedBlock implements a simple heap consisting of one contiguous area of memory organized in blocks, each of size theBlockSize. A block is a unit of allocation. A bitmap serves as bookkeeping data, more precisely one bit per block indicating whether that block is currently allocated or not.
Passing NullAllocator as ParentAllocator (the default) means user code manages allocation of the memory block from the outside; in that case BitmappedBlock must be constructed with a ubyte[] preallocated block and has no responsibility regarding the lifetime of its support underlying storage. If another allocator type is passed, BitmappedBlock defines a destructor that uses the parent allocator to release the memory block. That makes the combination of AllocatorList, BitmappedBlock, and a back-end allocator such as MmapAllocator a simple and scalable solution for memory allocation.
There are advantages to storing bookkeeping data separated from the payload (as opposed to e.g. using AffixAllocator to store metadata together with each allocation). The layout is more compact (overhead is one bit per block), searching for a free block during allocation enjoys better cache locality, and deallocation does not touch memory around the payload being deallocated (which is often cold).
Allocation requests are handled on a first-fit basis. Although linear in complexity, allocation is in practice fast because of the compact bookkeeping representation, use of simple and fast bitwise routines, and caching of the first available block position. A known issue with this general approach is fragmentation, partially mitigated by coalescing. Since BitmappedBlock does not need to maintain the allocated size, freeing memory implicitly coalesces free blocks together. Also, tuning blockSize has a considerable impact on both internal and external fragmentation.
If the last template parameter is set to No.multiblock, the allocator will only serve allocations which require at most theBlockSize. The BitmappedBlock has a specialized implementation for single-block allocations which allows for greater performance, at the cost of not being able to allocate more than one block at a time.
The size of each block can be selected either during compilation or at run time. Statically-known block sizes are frequent in practice and yield slightly better performance. To choose a block size statically, pass it as the blockSize parameter as in BitmappedBlock!(4096). To choose a block size parameter, use BitmappedBlock!(chooseAtRuntime) and pass the block size to the constructor.