C++: Custom memory allocation

Fast memory allocations along with memory leak detection can have a big impact on games performance.

C++ provides two well known functions to allocate dynamic (heap) memory (malloc and new), these functions are usually very slow because they're general purpose functions and in some implementations require a context-switch from user mode into kernel mode. These functions also do not provide any kind of memory leak detection system natively.

Using custom allocators we can have well defined usage patterns and optimize the allocation process accordingly.

Full source code and performance tests in the attached file

Base Allocator

Every allocator in this articles series will be derived from the class Allocator that declares 2 virtual functions (allocate and deallocate) that must be defined by each allocator.

Allocator.h

class Allocator{public:Allocator(){_usedMemory     = 0;_numAllocations = 0;}virtual ~Allocator(){ASSERT(_numAllocations == 0 && _usedMemory == 0);}virtual void* allocate(u32 size, u8 alignment) = 0;virtual void deallocate(void* p) = 0;template <class T> T* allocateNew(){return new (allocate(sizeof(T), __alignof(T))) T;}template <class T> T* allocateNew(const T& t){return new (allocate(sizeof(T), __alignof(T))) T(t);}template<class T> void deallocateDelete(T* pObject){if (pObject != nullptr){pObject->~T();deallocate(pObject);}}template<class T> T* allocateArray(u32 length){if(length == 0)return nullptr;u32 headerSize = sizeof(u32)/sizeof(T);if(sizeof(u32)%sizeof(T) > 0)headerSize += 1;//Allocate extra space to store array length in the bytes before the arrayT* p = ( (T*) allocate(sizeof(T)*(length + headerSize), __alignof(T)) ) + headerSize;*( ((u32*)p) - 1 ) = length;for(u32 i = 0; i < length; i++)new (&p[i]) T;return p;}template<class T> void deallocateArray(T* pArray){if(pArray == nullptr)return;u32 length = *( ((u32*)pArray) - 1 );for(int i = length-1; i >= 0; i--)pArray[i].~T();//Calculate how much extra memory was allocated to store the length before the arrayu32 headerSize = sizeof(u32)/sizeof(T);if(sizeof(u32)%sizeof(T) > 0)headerSize += 1;deallocate(pArray - headerSize);}u32 getUsedMemory(){return _usedMemory;}u32 getNumAllocations(){return _numAllocations;}protected:u32        _usedMemory;u32        _numAllocations;};

Memory leak detection

In the code above you can see an assert in the destructor, this is a simple and easy way to check if you forgot to deallocate any memory, that won't cause any overhead or take any extra memory.

This simple method won't tell which allocation you forgot to deallocate but it will pin point in which allocator the leak occured so you can find the leak faster (especially if you use Proxy Allocators like I suggest later in this article).

Aligned Allocations

Processors access memory in word-sized blocks, so when a processor tries to access memory in an unaligned address it might have to access more word-sized memory blocks than would be needed if the memory was aligned and perform masking/shifting to get the required data in the register.

Example:
A processor accesses memory in 4-byte words (it can only directly access the words starting at (0x00, 0x04, 0x08, 0x0C,...).

If it needs to access data (4 bytes) stored at the address 0x0B it will have to read two word-sized blocks (the address 0x08 and the address 0x0C) because the data crosses the boundary from one of the blocks to the other:



If the data was stored in an aligned address like 0x0C the processor could read it in a single memory access:

Aligned data definition

Primitive data is said to be aligned if the memory address where it is stored is a multiple of the size of the primitive.

A data aggregate is said to be aligned if each primitive element in the aggregate is aligned.

Implementation

To n-byte align a memory address x we need to mask off the log2(n) least significant bits from x. 

Simply masking off bits will return the first n-byte aligned address before x, so in order to find the first after x we just need to add alignment-1 to xand mask that address.

inline void* nextAlignedAddress(void* pAddress, u8 alignment){return (void*)( ( (uptr)pAddress + (alignment-1) ) & ~(alignment-1) );}

It can be useful to calculate by how many bytes the address needs to adjusted to be aligned.

inline u8 alignAdjustment(void* pAddress, u8 alignment){    u8 adjustment =  alignment - ( (uptr)pAddress & (alignment-1) );        if(adjustment == alignment)        return 0; //already aligned        return adjustment;}

Some allocators need to store an header before each allocation so they can use the adjustment space to reduce the memory overhead caused by the headers.

inline u8 alignAdjustmentWithHeader(void* pAddress, u8 alignment, u8 headerSize){    u8 adjustment =  alignment - ( (uptr)pAddress & (alignment-1) );        if(adjustment == alignment)        adjustment = 0; //already alignedu8 neededSpace = headerSize;if(adjustment < neededSpace){neededSpace -= adjustment;adjustment += alignment * (neededSpace / alignment);if(neededSpace % alignment > 0)adjustment += alignment;}        return adjustment;}

Note:  The alignment must be a power of 2!

Linear Allocator

A Linear Allocator is the simplest and fastest type of allocator. Pointers to the start of the allocator, to the first free address and the total size of the allocator are maintained. 

Allocations

New allocations simply move the pointer to the first free address forward.

Deallocations

Individual deallocations cannot be made in linear allocators, instead use clear() to completely clear the memory used by the allocator.

Implementation

LinearAllocator.h

#include "Allocator.h"#include "Types.h"class LinearAllocator : public Allocator{public:LinearAllocator(u32 size, void* pStart);~LinearAllocator();void* allocate(u32 size, u8 alignment);void deallocate(void* p);void clear();private:LinearAllocator(const LinearAllocator&) {}; //Prevent copies because it might cause errorsLinearAllocator& operator=(const LinearAllocator&) {};void* _pInitialPosition;void* _pCurrentPosition;u32   _size;};

LinearAllocator.cpp

#include "LinearAllocator.h"#include "Debug.h"LinearAllocator::LinearAllocator(u32 size, void* pStart) : Allocator(), _size(size), _pInitialPosition(pStart), _pCurrentPosition(pStart){ASSERT(size > 0);}LinearAllocator::~LinearAllocator(){_pInitialPosition   = nullptr;_pCurrentPosition   = nullptr;_size               = 0;}void* LinearAllocator::allocate(u32 size, u8 alignment){ASSERT(size != 0);u8 adjustment =  alignAdjustment(_pCurrentPosition, alignment);if(_usedMemory + adjustment + size > _size)return nullptr;uptr alignedAddress = (uptr)_pCurrentPosition + adjustment;_pCurrentPosition = (void*)(alignedAddress + size);_usedMemory += size + adjustment;_numAllocations++;return (void*)alignedAddress;}void LinearAllocator::deallocate(void* p){ASSERT( false && "Use clear() instead" );}void LinearAllocator::clear(){_numAllocations     = 0;_usedMemory         = 0;_pCurrentPosition   = _pInitialPosition;}

Stack Allocator

A Stack Allocator, like the name says, works like a stack. Along with the stack size, three pointers are maintained:

• Pointer to the start of the stack.
• Pointer to the top of the stack.
• Pointer to the last allocation made. (This is optional in release builds)

Allocations

New allocations move the pointer up by the requested number of bytes plus the adjustment needed to align the address and store the allocation header.

The allocation header provides the following information:

• Adjustment used in this allocation
• Pointer to the previous allocation.

Deallocations

Note:  Memory must be deallocated in inverse order it was allocated! So if you allocate object A and then object B you must free object B memory before you can free object A memory.

To deallocate memory the allocator checks if the address to the memory that you want to deallocate corresponds to the address of the last allocation made.

If so the allocator accesses the allocation header so it also frees the memory used to align the allocation and store the allocation header, and it replaces the pointer to the last allocation made with the one in the allocation header.

Implementation

StackAllocator.h

#include "Allocator.h"#include "Types.h"class StackAllocator : public Allocator{public:StackAllocator(u32 size, void* pStart);~StackAllocator();void* allocate(u32 size, u8 alignment);void deallocate(void* p);private:StackAllocator(const StackAllocator&) {}; //Prevent copies because it might cause errorsStackAllocator& operator=(const StackAllocator&) {};struct AllocationHeader{#if _DEBUGvoid* pPrevAddress;#endifu8 adjustment;};void* _pInitialPosition;#if _DEBUGvoid* _pPrevPosition;#endifvoid* _pCurrentPosition;u32   _size;};

StackAllocator.cpp

#include "StackAllocator.h"#include "Debug.h"StackAllocator::StackAllocator(u32 size, void* pStart) : Allocator(), _size(size), _pInitialPosition(pStart), _pCurrentPosition(pStart){ASSERT(size > 0);#if _DEBUG_pPrevPosition    = nullptr;#endif}StackAllocator::~StackAllocator(){_pInitialPosition   = nullptr;#if _DEBUG_pPrevPosition      = nullptr;#endif_pCurrentPosition   = nullptr;_size               = 0;}void* StackAllocator::allocate(u32 size, u8 alignment){ASSERT(size != 0);u8 adjustment = alignAdjustmentWithHeader(_pCurrentPosition, alignment, sizeof(AllocationHeader));if(_usedMemory + adjustment + size > _size)return nullptr;uptr alignedAddress = (uptr)_pCurrentPosition + adjustment;//Add Allocation HeaderAllocationHeader* pHeader = (AllocationHeader*)(alignedAddress-sizeof(AllocationHeader));pHeader->adjustment   = adjustment;#if _DEBUGpHeader->pPrevAddress = _pPrevPosition;_pPrevPosition    = (void*)alignedAddress;#endif_pCurrentPosition = (void*)(alignedAddress + size);_usedMemory += size + adjustment;_numAllocations++;return (void*)alignedAddress;}void StackAllocator::deallocate(void* p){ASSERT( p == _pPrevPosition );//Access the AllocationHeader in the bytes before pAllocationHeader* pHeader = (AllocationHeader*)((uptr)p - sizeof(AllocationHeader));_usedMemory -= (uptr)_pCurrentPosition - (uptr)p + pHeader->adjustment;_pCurrentPosition = (void*)( (uptr)p - pHeader->adjustment );#if _DEBUG_pPrevPosition = pHeader->pPrevAddress;#endif_numAllocations--;}

Note:  Storing the last previous allocations in a list-like fashion and checking it before deallocations is not mandatory so it can be disabled in release builds. It's just helpful to prevent memory from being overwritten causing bugs.

FreeList Allocator

The FreeList allocator allows allocations of any size to be made (inside the available memory) and deallocations in any order.

A linked-list of free blocks of memory is maintained (each free block contains information about its size and a pointer to the next free block).

Allocations

The allocator tries to find a free block large enough for the allocation to fit, if it finds multiple free blocks that meet the requeriments, there's 3 simple ways to decide which free block to choose:

• First-fit - Use the first.
• Best-fit - Use the smallest.
• Worst-fit - Use the largest.

Note:  The best-fit method will in most cases cause less fragmentation than the other 2 methods.

In the example implementation below I use the first-fit method.

Deallocation

The allocators recursively finds free blocks adjacent to the allocation being deallocated and merges them together.

If it doesn't finds any adjacent free blocks it creates a new free block and adds it to the list.

Implementation

FreeListAllocator.h

#include "Allocator.h"#include "Types.h"class FreeListAllocator : public Allocator{public:FreeListAllocator(u32 size, void* pStart);~FreeListAllocator();void* allocate(u32 size, u8 alignment);void deallocate(void* p);private:struct AllocationHeader{u32 size;u32 adjustment;};struct FreeBlock{u32 size;FreeBlock* pNext;};FreeListAllocator(const FreeListAllocator&) {}; //Prevent copies because it might cause errorsFreeListAllocator& operator=(const FreeListAllocator&) {};FreeBlock* _pFreeBlocks;};

FreeListAllocator.cpp

#include "FreeListAllocator.h"#include "Debug.h"FreeListAllocator::FreeListAllocator(u32 size, void* pStart) : Allocator(), _pFreeBlocks((FreeBlock*)pStart){ASSERT(size > sizeof(FreeBlock));_pFreeBlocks->size  = size;_pFreeBlocks->pNext = nullptr;}FreeListAllocator::~FreeListAllocator(){_pFreeBlocks        = nullptr;}void* FreeListAllocator::allocate(u32 size, u8 alignment){ASSERT(size != 0);//Check free blocksFreeBlock* pPrevFreeBlock = nullptr;FreeBlock* pFreeBlock     = _pFreeBlocks;while(pFreeBlock){//Calculate adjustment needed to keep object correctly alignedu8 adjustment = alignAdjustmentWithHeader(pFreeBlock, alignment, sizeof(AllocationHeader));//If allocation doesn't fit in this FreeBlock, try the nextif(pFreeBlock->size < size + adjustment){pPrevFreeBlock = pFreeBlock;pFreeBlock = pFreeBlock->pNext;continue;}ASSERT(sizeof(AllocationHeader) >= sizeof(FreeBlock));//If allocations in the remaining memory will be impossibleif(pFreeBlock->size - size - adjustment <= sizeof(AllocationHeader)){//Increase allocation size instead of creating a new FreeBlocksize = pFreeBlock->size;if(pPrevFreeBlock != nullptr)pPrevFreeBlock = pFreeBlock->pNext;else_pFreeBlocks = pFreeBlock->pNext;}else{//Else create a new FreeBlock containing remaining memoryFreeBlock* pNextBlock = (FreeBlock*)( (uptr)pFreeBlock + size + adjustment );pNextBlock->size = pFreeBlock->size - size - adjustment;pNextBlock->pNext = pFreeBlock->pNext;if(pPrevFreeBlock != nullptr)pPrevFreeBlock = pNextBlock;else_pFreeBlocks = pNextBlock;}uptr alignedAddress = (uptr)pFreeBlock + adjustment;AllocationHeader* pHeader = (AllocationHeader*)(alignedAddress - sizeof(AllocationHeader));pHeader->size             = size + adjustment;pHeader->adjustment       = adjustment;_usedMemory     += adjustment + size;_numAllocations++;return (void*)alignedAddress;}ASSERT(false && "Couldn't find free block large enough!");return nullptr;}void FreeListAllocator::deallocate(void* p){AllocationHeader* pHeader = (AllocationHeader*)((uptr)p - sizeof(AllocationHeader));u32 size = pHeader->size;uptr blockStart = (uptr)p - pHeader->adjustment;uptr blockEnd   = blockStart + size;u32 blockSize   = size;bool blockMerged = false;//Find adjacent free blocks and mergebool search = true;while(search){search = false;FreeBlock* pPrevFreeBlock = nullptr;FreeBlock* pFreeBlock = _pFreeBlocks;while(pFreeBlock != nullptr){if( (uptr)pFreeBlock + pFreeBlock->size == blockStart ){pFreeBlock->size += blockSize;blockStart = (uptr)pFreeBlock;blockEnd   = blockStart + pFreeBlock->size;blockSize  = pFreeBlock->size;search = true;blockMerged = true;break;} else if(blockEnd == (uptr) pFreeBlock){FreeBlock* pNewFreeBlock = (FreeBlock*) blockStart;pNewFreeBlock->pNext = pFreeBlock->pNext;pNewFreeBlock->size = blockSize + pFreeBlock->size;if(pFreeBlock == _pFreeBlocks)_pFreeBlocks = pNewFreeBlock;else if(pPrevFreeBlock != pNewFreeBlock)pPrevFreeBlock->pNext = pNewFreeBlock;blockStart = (uptr)pNewFreeBlock;blockEnd   = blockStart + pNewFreeBlock->size;blockSize  = pNewFreeBlock->size;search = true;blockMerged = true;break;}pPrevFreeBlock = pFreeBlock;pFreeBlock = pFreeBlock->pNext;}}if(!blockMerged){FreeBlock* pBlock         = (FreeBlock*)( (uptr)p - pHeader->adjustment);pBlock->size              = blockSize;pBlock->pNext             = _pFreeBlocks;_pFreeBlocks              = pBlock;}_numAllocations--;_usedMemory -= size;}

Pool Allocator

This allocator only allows allocations of a fixed size and alignment to be made, this results in both fast allocations and deallocations to be made. 

Like the FreeList allocator, a linked-list of free blocks is maintaied but since all blocks are the same size each free block only needs to store a pointer to the next one.

Another advantage of Pool allactors is no need to align each allocation, since all allocations have the same size/alignment only the first block has to be aligned, this results in a almost non-existant memory overhead.

Note:  
The block size of the Pool Allocator must be larger than sizeof(void*) because when blocks are free they store a pointer to the next free block in the list.

Allocations

The allocator simply returns the first free block and updates the linked list.

Deallocations

The allocator simply adds the deallocated block to the free blocks linked list.

Implementation

PoolAllocator.h

#include "Allocator.h"#include "Types.h"class PoolAllocator : public Allocator{public:PoolAllocator(u32 objectSize, u8 objectAlignment, u32 size, void* pMem);~PoolAllocator();void* allocate(u32 size, u8 alignment);void deallocate(void* p);private:PoolAllocator(const PoolAllocator&) {}; //Prevent copies because it might cause errorsPoolAllocator& operator=(const PoolAllocator&) {};u32        _size;u32        _objectSize;u8         _objectAlignment;void**     _pFreeList;};

PoolAllocator.cpp

#include "PoolAllocator.h"#include "Debug.h"PoolAllocator::PoolAllocator(u32 objectSize, u8 objectAlignment, u32 size, void* pMem) : Allocator(), _objectSize(objectSize), _objectAlignment(objectAlignment), _size(size){ASSERT(objectSize >= sizeof(void*));//Calculate adjustment needed to keep object correctly alignedu8 adjustment = alignAdjustment(pMem, objectAlignment);_pFreeList = (void**)((uptr)pMem + adjustment);u32 numObjects = (size-adjustment)/objectSize;void** p = _pFreeList;//Initialize free blocks listfor(u32 i = 0; i < numObjects-1; i++){*p = (void*)( (uptr) p + objectSize );p = (void**) *p;}*p = nullptr;}PoolAllocator::~PoolAllocator(){_pFreeList = nullptr;}void* PoolAllocator::allocate(u32 size, u8 alignment){ASSERT(size == _objectSize && alignment == _objectAlignment);if(_pFreeList == nullptr)return nullptr;void* p = _pFreeList;_pFreeList = (void**)(*_pFreeList);_usedMemory += size;_numAllocations++;return p;}void PoolAllocator::deallocate(void* p){*((void**)p) = _pFreeList;_pFreeList = (void**)p;_usedMemory -= _objectSize;_numAllocations--;}

Proxy Allocator

A Proxy Allocator is a special kind of allocator. It is just used to help with memory leak and subsystem memory usage tracking.

It will simply redirect all allocations/deallocations to the allocator passed as argument in the constructor while keeping track of how many allocations it made and how much memory it is "using".

Example:
Two subsystems use the same allocator A.
If you want to show in the debugging user interface how much memory each subsystem is using, you create a proxy allocator, that redirects all allocations/deallocations to A, in each subsystem and track their memory usage.

It will also help in memory leak tracking because the assert in the proxy allocator destructor of the subsystem that is leaking memory will fail.

Implementation

ProxyAllocator.h

#include "Allocator.h"class ProxyAllocator : public Allocator{public:ProxyAllocator(Allocator* pAllocator);    ~ProxyAllocator();        void* allocate(size_t size, size_t alignment);        void deallocate(void* p);    private:    ProxyAllocator(const ProxyAllocator&) {}; //Prevent copies because it might cause errors    ProxyAllocator& operator=(const ProxyAllocator&) {};        Allocator* _pAllocator;};

ProxyAllocator.cpp

#include "ProxyAllocator.h"#include "Debug.h"ProxyAllocator::ProxyAllocator(Allocator* pAllocator) : Allocator(), _pAllocator(pAllocator){ASSERT(pAllocator != NULL);}ProxyAllocator::~ProxyAllocator(){_pAllocator = nullptr;}void* ProxyAllocator::allocate(size_t size, size_t alignment){ASSERT(_pAllocator != NULL);_numAllocations++;u32 mem = _pAllocator->getUsedMemory();void* p = _pAllocator->allocate(size, alignment);_usedMemory += _pAllocator->getUsedMemory() - mem;return p;}void ProxyAllocator::deallocate(void* p){ASSERT(_pAllocator != NULL);_numAllocations--;u32 mem = _pAllocator->getUsedMemory();_pAllocator->deallocate(p);_usedMemory -= mem - _pAllocator->getUsedMemory();}

Allocator Managment

A large block of memory should be allocated when the program starts using malloc (and this should be the only malloc made) this large block of memory is managed by a global allocator (for example a stack allocator).

Each subsystem should then allocate the block of memory it needs to work from the global allocator, and create allocators that will manage that memory.

Example usage

• Allocate 1GB of memory using malloc and create a FreeList allocator to manage that memory.
• Create a Proxy allocator that redirects all allocations to the FreeList allocator.
• Initialize the Resource Manager by passing a pointer to the Proxy allocator in the constructor.
• Register the Proxy allocator in the memory usage tracker so it shows how much memory the Resource Manager is using.
• Allocate 16MB of memory using the FreeList allocator and create a Linear allocator to manage that memory and register it in the memory usage tracker.
• Use the Linear allocator to make small temporary allocations needed for game logic, etc, and clear it before the end of each frame.
• The Resource Manager will create a Pool allocator for every ResourcePackage it loads.

Tips & Tricks

• Depending on the type of allocator, keep the number of individual allocations to a minimum to reduce the memory wasted by allocation headers.
• Prefer using allocateArray() to individual allocations when it makes sense. Most allocators will use extra memory in each allocation to storeallocation headers and arrays will only need single header.
• Instead of making small size allocations from allocators with large amounts of memory available, allocate a single memory block capable of holding all the small allocations and create a new allocator to manage the memory block and make the small allocations from this block.

Performance Comparison

To test the performance of each allocator compared to malloc I wrote a program that measures how long it takes to make 20000 allocations (you can download the program in the end of the article), the tests where made in release mode and the results are averages of 3 runs.

Malloc vs Linear Allocator

10k 16 bytes allocations + 1k 256 bytes allocations + 50 2Mb allocations/deallocations (allocations made using the linear allocator are deallocated in a single call to clear().
AllocatorTime (s)Malloc0.639655Linear0.000072

Malloc vs Stack Allocator

10k 16 bytes allocations + 1k 256 bytes allocations + 50 2Mb allocations/deallocations
AllocatorTime (s)Malloc0.650435Stack0.000289

Malloc vs FreeList Allocator

10k 16 bytes allocations + 1k 256 bytes allocations + 50 2Mb allocations/deallocations
AllocatorTime (s)Malloc0.673865FreeList0.000414

Malloc vs Pool Allocator

20k 16 bytes allocations/deallocations
AllocatorTime (s)Malloc1.454934Pool0.000193

Conclusion

There isn't a single best allocator - it's important to think about how the memory will be allocated/accessed/deallocated and choose the right allocator for each situation.

Full source code and performance tests in the attached file

Reference

http://bitsquid.blogspot.pt/2010/09/custom-memory-allocation-in-c.html
Game Engine Architecture, Jason Gregory 2009
http://molecularmusings.wordpress.com/2011/07/05/memory-system-part-1/

Article Update Log

14 April 2013: Fixed an error in allocateArray
13 April 2013: FreeList and Pool allocators added
12 April 2013: Performance comparison & test project added
11 April 2013: Initial release

 Download attached article resource

About the Author(s)

My name is Tiago Costa and I'm currently studying Computer Science at FCUP, Portugal and working on an indie project as a graphics/game programmer. 

License

GDOL (Gamedev.net Open License)

转自 http://www.gamedev.net/page/resources/_/technical/general-programming/c-custom-memory-allocation-r3010