Report a bug
If you spot a problem with this page, click here to create a Bugzilla issue.
Improve this page
Quickly fork, edit online, and submit a pull request for this page. Requires a signed-in GitHub account. This works well for small changes. If you'd like to make larger changes you may want to consider using a local clone.

std.experimental.allocator.building_blocks.kernighan_ritchie

struct KRRegion(ParentAllocator = NullAllocator);
KRRegion draws inspiration from the region allocation strategy and also the famed allocator described by Brian Kernighan and Dennis Ritchie in section 8.7 of the book "The C Programming Language", Second Edition, Prentice Hall, 1988.

KRRegion = Region + Kernighan-Ritchie Allocator

Initially, KRRegion starts in "region" mode: allocations are served from the memory chunk in a region fashion. Thus, as long as there is enough memory left, KRRegion.allocate has the performance profile of a region allocator. Deallocation inserts (in Ο(1) time) the deallocated blocks in an unstructured freelist, which is not read in region mode.
Once the region cannot serve an allocate request, KRRegion switches to "free list" mode. It sorts the list of previously deallocated blocks by address and serves allocation requests off that free list. The allocation and deallocation follow the pattern described by Kernighan and Ritchie.
The recommended use of KRRegion is as a region with deallocation. If the KRRegion is dimensioned appropriately, it could often not enter free list mode during its lifetime. Thus it is as fast as a simple region, whilst offering deallocation at a small cost. When the region memory is exhausted, the previously deallocated memory is still usable, at a performance cost. If the region is not excessively large and fragmented, the linear allocation and deallocation cost may still be compensated for by the good locality characteristics.
If the chunk of memory managed is large, it may be desirable to switch management to free list from the beginning. That way, memory may be used in a more compact manner than region mode. To force free list mode, call switchToFreeList shortly after construction or when deemed appropriate.
The smallest size that can be allocated is two words (16 bytes on 64-bit systems, 8 bytes on 32-bit systems). This is because the free list management needs two words (one for the length, the other for the next pointer in the singly-linked list).
The ParentAllocator type parameter is the type of the allocator used to allocate the memory chunk underlying the KRRegion object. Choosing the default (NullAllocator) means the user is responsible for passing a buffer at construction (and for deallocating it if necessary). Otherwise, KRRegion automatically deallocates the buffer during destruction. For that reason, if ParentAllocator is not NullAllocator, then KRRegion is not copyable.

Implementation Details

In free list mode, KRRegion embeds a free blocks list onto the chunk of memory. The free list is circular, coalesced, and sorted by address at all times. Allocations and deallocations take time proportional to the number of previously deallocated blocks. (In practice the cost may be lower, e.g. if memory is deallocated in reverse order of allocation, all operations take constant time.) Memory utilization is good (small control structure and no per-allocation overhead). The disadvantages of freelist mode include proneness to fragmentation, a minimum allocation size of two words, and linear worst-case allocation and deallocation times.
Similarities of KRRegion (in free list mode) with the Kernighan-Ritchie allocator:
  • Free blocks have variable size and are linked in a singly-linked list.
  • The freelist is maintained in increasing address order, which makes coalescing easy.
  • The strategy for finding the next available block is first fit.
  • The free list is circular, with the last node pointing back to the first.
  • Coalescing is carried during deallocation.
Differences from the Kernighan-Ritchie allocator:
  • Once the chunk is exhausted, the Kernighan-Ritchie allocator allocates another chunk using operating system primitives. For better composability, KRRegion just gets full (returns null on new allocation requests). The decision to allocate more blocks is deferred to a higher-level entity. For an example, see the example below using AllocatorList in conjunction with KRRegion.
  • Allocated blocks do not hold a size prefix. This is because in D the size information is available in client code at deallocation time.
Examples:
KRRegion is preferable to Region as a front for a general-purpose allocator if deallocate is needed, yet the actual deallocation traffic is relatively low. The example below shows a KRRegion using stack storage fronting the GC allocator.
import std.experimental.allocator.building_blocks.fallback_allocator
    : fallbackAllocator;
import std.experimental.allocator.gc_allocator : GCAllocator;
import std.typecons : Ternary;
// KRRegion fronting a general-purpose allocator
ubyte[1024 * 128] buf;
auto alloc = fallbackAllocator(KRRegion!()(buf), GCAllocator.instance);
auto b = alloc.allocate(100);
writeln(b.length); // 100
writeln((() pure nothrow @safe @nogc => alloc.primary.owns(b))()); // Ternary.yes
Examples:
The code below defines a scalable allocator consisting of 1 MB (or larger) blocks fetched from the garbage-collected heap. Each block is organized as a KR-style heap. More blocks are allocated and freed on a need basis.
This is the closest example to the allocator introduced in the K&R book. It should perform slightly better because instead of searching through one large free list, it searches through several shorter lists in LRU order. Also, it actually returns memory to the operating system when possible.
import std.algorithm.comparison : max;
import std.experimental.allocator.building_blocks.allocator_list
    : AllocatorList;
import std.experimental.allocator.mmap_allocator : MmapAllocator;
AllocatorList!(n => KRRegion!MmapAllocator(max(n * 16, 1024 * 1024))) alloc;
ParentAllocator parent;
If ParentAllocator holds state, parent is a public member of type KRRegion. Otherwise, parent is an alias for ParentAllocator.instance.
this(ubyte[] b);

this(size_t n);

this(ParentAllocator parent, size_t n);
Create a KRRegion. If ParentAllocator is not NullAllocator, KRRegion's destructor will call parent.deallocate.
Parameters:
ubyte[] b Block of memory to serve as support for the allocator. Memory must be larger than two words and word-aligned.
size_t n Capacity desired. This constructor is defined only if ParentAllocator is not NullAllocator.
void switchToFreeList();
Forces free list mode. If already in free list mode, does nothing. Otherwise, sorts the free list accumulated so far and switches strategy for future allocations to KR style.
enum auto alignment;
Word-level alignment.
void[] allocate(size_t n);
Allocates n bytes. Allocation searches the list of available blocks until a free block with n or more bytes is found (first fit strategy). The block is split (if larger) and returned.
Parameters:
size_t n number of bytes to allocate
Returns:
A word-aligned buffer of n bytes, or null.
nothrow @nogc bool deallocate(void[] b);
Deallocates b, which is assumed to have been previously allocated with this allocator. Deallocation performs a linear search in the free list to preserve its sorting order. It follows that blocks with higher addresses in allocators with many free blocks are slower to deallocate.
Parameters:
void[] b block to be deallocated
void[] allocateAll();
Allocates all memory available to this allocator. If the allocator is empty, returns the entire available block of memory. Otherwise, it still performs a best-effort allocation: if there is no fragmentation (e.g. allocate has been used but not deallocate), allocates and returns the only available block of memory.
The operation takes time proportional to the number of adjacent free blocks at the front of the free list. These blocks get coalesced, whether allocateAll succeeds or fails due to fragmentation.
Examples:
import std.experimental.allocator.gc_allocator : GCAllocator;
auto alloc = KRRegion!GCAllocator(1024 * 64);
const b1 = alloc.allocate(2048);
writeln(b1.length); // 2048
const b2 = alloc.allocateAll;
writeln(b2.length); // 1024 * 62
pure nothrow @nogc bool deallocateAll();
Deallocates all memory currently allocated, making the allocator ready for other allocations. This is a Ο(1) operation.
pure nothrow @nogc @trusted Ternary owns(void[] b);
Checks whether the allocator is responsible for the allocation of b. It does a simple Ο(1) range check. b should be a buffer either allocated with this or obtained through other means.
static pure nothrow @nogc @safe size_t goodAllocSize(size_t n);
Adjusts n to a size suitable for allocation (two words or larger, word-aligned).
pure nothrow @nogc @safe Ternary empty();
Returns:
Ternary.yes if the allocator is empty, Ternary.no otherwise. Never returns Ternary.unknown.