#include "heap.h"

#include <algorithm>									// lower_bound, min
#include <cstring>										// strlen, memset, memcpy
#include <climits>										// ULONG_MAX
#include <cstdarg>										// va_start, va_end
#include <cerrno>										// errno, ENOMEM, EINVAL
#include <cassert>
#include <unistd.h>										// STDERR_FILENO, sbrk, sysconf, write
#include <sys/mman.h>									// mmap, munmap
#include <cstdint>										// uintptr_t, uint64_t, uint32_t

#define TLS
#define AWAYSPIN										// toggle spinlock / lockfree stack
#define FASTLOOKUP										// use O(1) table lookup from allocation size to bucket size
#define CFA_THREADSAFE_HEAP

#ifdef __CFA_DEBUG__
#define __DEBUG__
#endif // __CFA_DEBUG__

#define LIKELY(x) __builtin_expect(!!(x), 1)
#define UNLIKELY(x) __builtin_expect(!!(x), 0)

#define str(s) #s
#define xstr(s) str(s)
#define WARNING( s ) xstr( GCC diagnostic ignored str( -W ## s ) )
#define NOWARNING( statement, warning ) \
	_Pragma( "GCC diagnostic push" ) \
	_Pragma( WARNING( warning ) ) \
	statement ;	\
	_Pragma ( "GCC diagnostic pop" )

enum { __ALIGN__ = 16,									// minimum allocation alignment, bytes
	   __DEFAULT_HEAP_EXPANSION__ = 2 * 1024 * 1024,	// sbrk extension amount when full
	   __DEFAULT_MMAP_START__ = 512 * 1024 + 1,			// crossover allocation size from sbrk to mmap
	   __DEFAULT_HEAP_UNFREED__ = 0,					// amount subtracted to adjust for unfreed program storage
}; // enum

#ifdef TLS
#define TLSMODEL __attribute__(( tls_model("initial-exec") ))
#else
#define TLSMODEL
#endif // TLS

#ifdef CFA_THREADSAFE_HEAP
extern "C" {
	void enable_interrupts();
	void disable_interrupts();
}
#define ENABLE_INTERRUPTS (void)enable_interrupts()
#define DISABLE_INTERRUPTS (void)disable_interrupts()
#else
#define ENABLE_INTERRUPTS
#define DISABLE_INTERRUPTS
#endif // CFA_THREADSAFE_HEAP

//######################### Helpers #########################


// Called by macro assert in assert.h. Replace to prevent recursive call to malloc.
/*
void __assert_fail( const char assertion[], const char file[], unsigned int line, const char function[] ) {
	extern const char * __progname;						// global name of running executable (argv[0])
	char helpText[1024];
	int len = snprintf( helpText, sizeof(helpText), "Internal assertion error \"%s\" from program \"%s\" in \"%s\" at line %d in file \"%s.\n",
						assertion, __progname, function, line, file );
	NOWARNING( write( STDERR_FILENO, helpText, len ), unused-result );
	abort();
	// CONTROL NEVER REACHES HERE!
} // __assert_fail
FIXME */

void abort( const char fmt[], ... ) __attribute__(( format(printf, 1, 2), __nothrow__, __leaf__, __noreturn__ ));
void abort( const char fmt[], ... ) {					// overload real abort
	va_list args;
	va_start( args, fmt );
	vfprintf( stderr, fmt, args );
	if ( fmt[strlen( fmt ) - 1] != '\n' ) {				// add optional newline if missing at the end of the format text
		vfprintf( stderr, "\n", args );					// g++-10 does not allow nullptr for va_list
	} // if
	va_end( args );
	abort();											// call the real abort
	// CONTROL NEVER REACHES HERE!
} // abort

static inline bool Pow2( unsigned long int value ) {
	// clears all bits below value, rounding value down to the next lower multiple of value
	return (value & (value - 1)) == 0;
} // Pow2

static inline unsigned long int Floor( unsigned long int value, unsigned long int align ) {
	assert( Pow2( align ) );
	// clears all bits above or equal to align, getting (value % align), the phase of value with regards to align
	return value & -align;
} // Floor

static inline unsigned long int Ceiling( unsigned long int value, unsigned long int align ) {
	assert( Pow2( align ) );
	// "negate, round down, negate" is the same as round up
	return -Floor( -value, align );
} // Ceiling

template< typename T > static inline T AtomicFetchAdd( volatile T & counter, int increment ) {
	return __atomic_fetch_add( &counter, increment, __ATOMIC_SEQ_CST );
} // AtomicFetchAdd


//######################### Spin Lock #########################


#define CACHE_ALIGN 128									// Intel recommendation
#define CALIGN __attribute__(( aligned(CACHE_ALIGN) ))

// pause to prevent excess processor bus usage
#if defined( __i386 ) || defined( __x86_64 )
	#define Pause() __asm__ __volatile__ ( "pause" : : : )
#elif defined(__ARM_ARCH)
	#define Pause() __asm__ __volatile__ ( "YIELD" : : : )
#else
	#error unsupported architecture
#endif

typedef volatile uintptr_t SpinLock_t CALIGN;			// aligned addressable word-size

void spin_acquire( volatile SpinLock_t * lock ) {
	enum { SPIN_START = 4, SPIN_END = 64 * 1024, };
	unsigned int spin = SPIN_START;

	for ( unsigned int i = 1;; i += 1 ) {
	  if ( *lock == 0 && __atomic_test_and_set( lock, __ATOMIC_SEQ_CST ) == 0 ) break; // Fence
		for ( volatile unsigned int s = 0; s < spin; s += 1 ) Pause(); // exponential spin
		spin += spin;									// powers of 2
		//if ( i % 64 == 0 ) spin += spin;				// slowly increase by powers of 2
		if ( spin > SPIN_END ) spin = SPIN_END;			// cap spinning
	} // for
} // spin_lock

void spin_release( volatile SpinLock_t * lock ) {
	__atomic_clear( lock, __ATOMIC_SEQ_CST );			// Fence
} // spin_unlock


//####################### Heap Statistics ####################


#ifdef __STATISTICS__
enum { CntTriples = 12 };								// number of counter triples
struct HeapStatistics {
	enum { MALLOC, AALLOC, CALLOC, MEMALIGN, AMEMALIGN, CMEMALIGN, RESIZE, REALLOC };
	union {
		struct {
			unsigned int malloc_calls, malloc_0_calls;
			unsigned long long int malloc_storage_request, malloc_storage_alloc;
			unsigned int aalloc_calls, aalloc_0_calls;
			unsigned long long int aalloc_storage_request, aalloc_storage_alloc;
			unsigned int calloc_calls, calloc_0_calls;
			unsigned long long int calloc_storage_request, calloc_storage_alloc;
			unsigned int memalign_calls, memalign_0_calls;
			unsigned long long int memalign_storage_request, memalign_storage_alloc;
			unsigned int amemalign_calls, amemalign_0_calls;
			unsigned long long int amemalign_storage_request, amemalign_storage_alloc;
			unsigned int cmemalign_calls, cmemalign_0_calls;
			unsigned long long int cmemalign_storage_request, cmemalign_storage_alloc;
			unsigned int resize_calls, resize_0_calls;
			unsigned long long int resize_storage_request, resize_storage_alloc;
			unsigned int realloc_calls, realloc_0_calls;
			unsigned long long int realloc_storage_request, realloc_storage_alloc;
			unsigned int free_calls, free_null_calls;
			unsigned long long int free_storage_request, free_storage_alloc;
			unsigned int away_pulls, away_pushes;
			unsigned long long int away_storage_request, away_storage_alloc;
			unsigned int mmap_calls, mmap_0_calls;		// no zero calls
			unsigned long long int mmap_storage_request, mmap_storage_alloc;
			unsigned int munmap_calls, munmap_0_calls;	// no zero calls
			unsigned long long int munmap_storage_request, munmap_storage_alloc;
		};
		struct {										// overlay for iteration
			unsigned int cnt1, cnt2;
			unsigned long long int cnt3, cnt4;
		} counters[CntTriples];
	};

	HeapStatistics() {
		for ( unsigned int i = 0; i < CntTriples; i += 1 ) {
			counters[i].cnt1 = counters[i].cnt2 = counters[i].cnt3 = counters[i].cnt4 = 0;
		} // for
	} // HeapStatistics::HeapStatistics

	friend HeapStatistics & operator+=( HeapStatistics & lhs, const HeapStatistics & rhs ) {
		for ( unsigned int i = 0; i < CntTriples; i += 1 ) {
			lhs.counters[i].cnt1 += rhs.counters[i].cnt1;
			lhs.counters[i].cnt2 += rhs.counters[i].cnt2;
			lhs.counters[i].cnt3 += rhs.counters[i].cnt3;
			lhs.counters[i].cnt4 += rhs.counters[i].cnt4;
		} // for
		return lhs;
	} // HeapStatistics::operator+=
}; // HeapStatistics

static_assert( sizeof(HeapStatistics) == CntTriples * sizeof(HeapStatistics::counters[0] ),
			   "Heap statistics counter-triplets does not match with array size" );
#endif // __STATISTICS__


//####################### Heap Structure ####################


struct Heap {
	struct FreeHeader;									// forward declaration

	struct Storage {
		struct Header {									// header
			union Kind {
				struct RealHeader {
					union {
						struct {						// 4-byte word => 8-byte header, 8-byte word => 16-byte header
							#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ && __SIZEOF_POINTER__ == 4
							uint64_t padding;			// unused, force home/blocksize to overlay alignment in fake header
							#endif // __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ && __SIZEOF_POINTER__ == 4

							union {
								// 2nd low-order bit => zero filled, 3rd low-order bit => mmapped
								FreeHeader * home;		// allocated block points back to home locations (must overlay alignment)
								size_t blockSize;		// size for munmap (must overlay alignment)
								Storage * next;			// freed block points to next freed block of same size
							};
							size_t size;				// allocation size in bytes

							#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ && __SIZEOF_POINTER__ == 4
							uint64_t padding;			// unused, force home/blocksize to overlay alignment in fake header
							#endif // __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ && __SIZEOF_POINTER__ == 4
						};
					};
				} real; // RealHeader
				struct FakeHeader {
					#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
					uint32_t alignment;					// 1st low-order bit => fake header & alignment
					#endif // __ORDER_LITTLE_ENDIAN__

					uint32_t offset;

					#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
					uint32_t alignment;					// 1st low-order bit => fake header & alignment
					#endif // __ORDER_BIG_ENDIAN__
				} fake; // FakeHeader
			} kind; // Kind
		} header; // Header

		char pad[__ALIGN__ - sizeof( Header )];
		char data[0];									// storage
	}; // Storage

	static_assert( __ALIGN__ >= sizeof( Storage ), "minimum alignment < sizeof( Storage )" );

	struct FreeHeader {
		#ifdef AWAYSPIN
		SpinLock_t awayLock;							// LOCK(S) MUST BE FIRST FIELD(S) FOR ALIGNMENT
		#endif // AWAYSPIN

		Storage * freeList;								// thread free list
		Storage * awayList;								// other thread return list

		Heap * homeManager;								// heap owner (free storage to bucket, from bucket to heap)
		size_t blockSize;								// size of allocations on this list

		bool operator<( const size_t bsize ) const { return blockSize < bsize; }
	}; // FreeHeader

	// Recursive definitions: HeapManager needs size of bucket array and bucket area needs sizeof HeapManager storage.
	// Break recursion by hardcoding number of buckets and statically checking number is correct after bucket array defined.
	enum {
		#ifdef FASTLOOKUP
		LookupSizes = 65'536 + sizeof(Storage),			// number of fast lookup sizes '
		#endif // FASTLOOKUP
		NoBucketSizes = 91,								// number of bucket sizes
	}; // enum

	FreeHeader freeLists[NoBucketSizes];				// buckets for different allocation sizes
	void * heapBuffer;
	size_t heapReserve;

	Heap * nextHeapManager;								// intrusive link of existing heaps; traversed to collect statistics
	Heap * nextFreeHeapManager;							// intrusive link of free heaps from terminated threads; reused by new threads

	#ifdef __DEBUG__
	long long int allocUnfreed;							// running total of allocations minus frees; can be negative
	#endif // __DEBUG__

	#ifdef __STATISTICS__
	HeapStatistics stats;								// local statistic table for this heap
	#endif // __STATISTICS__

	static void heapManagerCtor(
		#ifdef __DEBUG__
		size_t size
		#endif // __DEBUG__
	);
	static void heapManagerDtor();
}; // Heap


struct ThreadManager {
	int dummy;											// used to trigger allocation of storage
	~ThreadManager() { Heap::heapManagerDtor(); }		// called automagically when thread terminates
}; // ThreadManager


struct HeapMaster {
	SpinLock_t masterExtLock;							// protects allocation-buffer extension
	SpinLock_t masterMgrLock;							// protects freeHeapManagersList, heapManagersList, heapManagersStorage, heapManagersStorageEnd

	#ifdef FASTLOOKUP
	unsigned char lookup[Heap::LookupSizes];			// O(1) lookup for small sizes
	#endif // FASTLOOKUP

	static const unsigned int bucketSizes[];			// initialized statically, outside constructor
	void * heapBegin;									// start of heap
	void * heapEnd;										// logical end of heap
	size_t heapRemaining;								// amount of storage not allocated in the current chunk
	size_t pageSize;									// architecture pagesize
	size_t heapExpand;									// sbrk advance
	size_t mmapStart;									// cross over point for mmap
	unsigned int maxBucketsUsed;						// maximum number of buckets in use
	static const off_t mmapFd;							// fake or actual fd for anonymous file. initialized statically, outside constructor

	Heap * heapManagersList;							// heap-list head
	Heap * freeHeapManagersList;						// free-list head

	// Heap superblocks are not linked; heaps in superblocks are linked via intrusive links.
	Heap * heapManagersStorage;							// next heap to use in heap superblock
	Heap * heapManagersStorageEnd;						// logical heap outside of superblock's end

	#ifdef __STATISTICS__
	unsigned long int threads_started, threads_exited;  // counts threads that have started and exited
	unsigned long int reused_heap, new_heap;  			// counts reusability of heaps
	unsigned int sbrk_calls;
	unsigned long long int sbrk_storage;
	int stats_fd;
	HeapStatistics stats;								// global stats for thread-local heaps to add there counters when exiting
	#endif // __STATISTICS__

	// Prevents two threads from constructing heapMaster.
	static volatile bool heapMasterBootFlag;			// trigger for first heap

	#ifdef __DEBUG__
	long long int allocUnfreed;
	#endif // __DEBUG__

	static void heapMasterCtor();
	static void heapMasterDtor();
}; // HeapMaster

extern "C" {
	int __map_prot = PROT_READ | PROT_WRITE | PROT_EXEC; 	// common mmap/mprotect protection
	size_t __cfa_page_size;									// architecture pagesize
}

volatile bool HeapMaster::heapMasterBootFlag = false;
static HeapMaster heapMaster;							// program global

// Thread-local storage is allocated lazily when the storage is accessed.
static thread_local size_t PAD1 CALIGN TLSMODEL __attribute__(( unused )); // protect false sharing
static thread_local ThreadManager threadManager CALIGN TLSMODEL;
// Do not put heapManager in ThreadManager because thread-local destructor results in extra access code.
static thread_local Heap * heapManager CALIGN TLSMODEL;
static thread_local bool heapManagerBootFlag CALIGN TLSMODEL = false;
static thread_local size_t PAD2 CALIGN TLSMODEL __attribute__(( unused )); // protect further false sharing


#ifdef __DEBUG__
extern "C" {
	void heapAppStart( void ) {
		assert( heapManager );
		heapManager->allocUnfreed = 0;
	} // heapAppStart

	void heapAppStop( void ) {
		long long int allocUnfreed = heapMaster.allocUnfreed;
		for ( Heap * heap = heapMaster.heapManagersList; heap; heap = heap->nextHeapManager ) {
			allocUnfreed += heap->allocUnfreed;
		} // for

		allocUnfreed -= malloc_unfreed();
		if ( allocUnfreed > 0 ) {
			// DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
			char helpText[512];
			int len = snprintf( helpText, sizeof(helpText), "Runtime warning (UNIX pid:%ld) : program terminating with %llu(0x%llx) bytes of storage allocated but not freed.\n"
								"Possible cause is unfreed storage allocated by the program or system/library routines called from the program.\n",
								(long int)getpid(), allocUnfreed, allocUnfreed ); // always print the UNIX pid
			NOWARNING( write( STDERR_FILENO, helpText, len ), unused-result );
		} // if
	} // heapAppStop
} // extern "C"
#endif // __DEBUG__


// declare helper functions for HeapMaster
void noMemory();										// forward, called by "builtin_new" when malloc returns 0

void HeapMaster::heapMasterCtor() {
	// Singleton pattern to initialize heap master
	__cfa_page_size = sysconf( _SC_PAGESIZE );

	assert( heapMaster.mmapFd == -1 );
	assert( heapMaster.bucketSizes[0] == (16 + sizeof(Heap::Storage)) );

	heapMaster.masterExtLock = 0;
	heapMaster.masterMgrLock = 0;

	char * end = (char *)sbrk( 0 );
	heapMaster.heapBegin = heapMaster.heapEnd = sbrk( (char *)Ceiling( (long unsigned int)end, __ALIGN__ ) - end ); // move start of heap to multiple of alignment
	heapMaster.heapRemaining = 0;
	heapMaster.heapExpand = malloc_expansion();
	heapMaster.mmapStart = malloc_mmap_start();

	// find the closest bucket size less than or equal to the mmapStart size
	heapMaster.maxBucketsUsed = std::lower_bound( heapMaster.bucketSizes, heapMaster.bucketSizes + (Heap::NoBucketSizes - 1), heapMaster.mmapStart ) - heapMaster.bucketSizes; // binary search

	assert( (heapMaster.mmapStart >= __cfa_page_size) && (heapMaster.bucketSizes[Heap::NoBucketSizes - 1] >= heapMaster.mmapStart) );
	assert( heapMaster.maxBucketsUsed < Heap::NoBucketSizes ); // subscript failure ?
	assert( heapMaster.mmapStart <= heapMaster.bucketSizes[heapMaster.maxBucketsUsed] ); // search failure ?

	heapMaster.heapManagersList = nullptr;
	heapMaster.freeHeapManagersList = nullptr;

	heapMaster.heapManagersStorage = nullptr;
	heapMaster.heapManagersStorageEnd = nullptr;

	#ifdef __STATISTICS__
	heapMaster.threads_started = heapMaster.threads_exited = 0;
	heapMaster.reused_heap = heapMaster.new_heap = 0;
	heapMaster.sbrk_calls = heapMaster.sbrk_storage = 0;
	heapMaster.stats_fd = STDERR_FILENO;
	#endif // __STATISTICS__

	#ifdef __DEBUG__
	heapMaster.allocUnfreed = 0;
	#endif // __DEBUG__

	#ifdef FASTLOOKUP
	for ( unsigned int i = 0, idx = 0; i < Heap::LookupSizes; i += 1 ) {
		if ( i > heapMaster.bucketSizes[idx] ) idx += 1;
		heapMaster.lookup[i] = idx;
		assert( i <= heapMaster.bucketSizes[idx] );
		assert( (i <= 32 && idx == 0) || (i > heapMaster.bucketSizes[idx - 1]) );
	} // for
	#endif // FASTLOOKUP

	std::set_new_handler( noMemory );					// do not throw exception as the default

	HeapMaster::heapMasterBootFlag = true;
} // HeapMaster::heapMasterCtor


#define NO_MEMORY_MSG "insufficient heap memory available for allocating %zd new bytes."

void Heap::heapManagerCtor(
		#ifdef __DEBUG__
		size_t size
		#endif // __DEBUG__
) {
	if ( UNLIKELY( ! HeapMaster::heapMasterBootFlag ) ) HeapMaster::heapMasterCtor();

	// Trigger thread_local storage implicit allocation (causes recursive call)
	volatile int dummy __attribute__(( unused )) = threadManager.dummy;

	spin_acquire( &heapMaster.masterMgrLock );			// protect heapMaster counters
	// The atomic test-and-set instruction is a fence so heapManagerBootFlag is read after the magic recursive call to
	// initialize thread-local storage. Hence, heapManagerBootFlag is NOT declared as volatile.
  if ( heapManagerBootFlag ) {							// singleton
		spin_release( &heapMaster.masterMgrLock );
		return;											// always return on recursive initiation
	} // if

	assert( ! heapManagerBootFlag );

	// get storage for heap manager

	if ( heapMaster.freeHeapManagersList ) {			// free heap for reused ?
		heapManager = heapMaster.freeHeapManagersList;
		heapMaster.freeHeapManagersList = heapManager->nextFreeHeapManager;

		#ifdef __STATISTICS__
		heapMaster.reused_heap += 1;
		#endif // __STATISTICS__
	} else {											// free heap not found, create new
		// Heap size is about 12K, FreeHeader (128 bytes because of cache alignment) * NoBucketSizes (91) => 128 heaps * 12K ~= 120K byte superblock.
		// Where 128-heap superblock handles a medium sized multi-processor server.
		enum { HeapDim = 128 };							// number of heaps in superblock
		size_t remaining = heapMaster.heapManagersStorageEnd - heapMaster.heapManagersStorage; // remaining free heaps in superblock
		if ( ! heapMaster.heapManagersStorage || remaining != 0 ) {
			size_t size = HeapDim * sizeof( Heap );
			heapMaster.heapManagersStorage = (Heap *)mmap( 0, size, __map_prot, MAP_PRIVATE | MAP_ANONYMOUS, heapMaster.mmapFd, 0 );
			if ( UNLIKELY( heapMaster.heapManagersStorage == MAP_FAILED ) ) { // failed ?
				if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, size ); // no memory
				// Do not call strerror( errno ) as it may call malloc.
				abort( "heapManagerCtor() : internal error, mmap failure, size:%zu error %d.",
					   size, errno );
			} // if
			heapMaster.heapManagersStorageEnd = &heapMaster.heapManagersStorage[HeapDim]; // outside array
		} // if

		heapManager = heapMaster.heapManagersStorage;
		heapMaster.heapManagersStorage = heapMaster.heapManagersStorage + 1; // bump next heap

		heapManager->nextHeapManager = heapMaster.heapManagersList;
		heapMaster.heapManagersList = heapManager;

		#ifdef __STATISTICS__
		heapMaster.new_heap += 1;
		#endif // __STATISTICS__
	} // if

	#ifdef __STATISTICS__
	heapMaster.threads_started += 1;
	#endif // __STATISTICS__

	#ifdef __DEBUG__
	heapManager->allocUnfreed -= size;
	#endif // __DEBUG__

	spin_release( &heapMaster.masterMgrLock );

	for ( unsigned int j = 0; j < Heap::NoBucketSizes; j += 1 ) { // initialize free lists
		heapManager->freeLists[j] = (Heap::FreeHeader){
			#ifdef AWAYSPIN
			.awayLock = 0,
			#endif // AWAYSPIN
			.freeList = nullptr,
			.awayList = nullptr,
			.homeManager = heapManager,
			.blockSize = heapMaster.bucketSizes[j],
		};
	} // for

	heapManager->heapBuffer = nullptr;
	heapManager->heapReserve = 0;
	heapManager->nextFreeHeapManager = nullptr;
	heapManagerBootFlag = true;
} // Heap::heapManagerCtor


void Heap::heapManagerDtor() {
  if ( UNLIKELY( ! heapManagerBootFlag ) ) return;

	DISABLE_INTERRUPTS;

	spin_acquire( &heapMaster.masterMgrLock );

	// place heap on list of free heaps for reusability
	heapManager->nextFreeHeapManager = heapMaster.freeHeapManagersList;
	heapMaster.freeHeapManagersList = heapManager;

	// SKULLDUGGERY: The thread heap ends BEFORE the last free(s) occurs from the thread-local storage allocations for
	// the thread. This final allocation must be handled in doFree for this thread and its terminated heap. However,
	// this heap has just been put on the heap freelist, and hence there is a race returning the thread-local storage
	// and a new thread using this heap. The current thread detects it is executing its last free in doFree via
	// heapManager being null. The trick is for this thread to placed the last free onto the current heap's away-list as
	// the free-storage header points are this heap. Now, even if other threads are pushing to the away list, it is safe
	// because of the locking.
	heapManager = nullptr;

	#ifdef __STATISTICS__
	heapMaster.threads_exited += 1;
	#endif // __STATISTICS__

	spin_release( &heapMaster.masterMgrLock );

	ENABLE_INTERRUPTS;
} // Heap::heapManagerDtor


const off_t HeapMaster::mmapFd = -1;
// Bucket size must be multiple of 16.
// Powers of 2 are common allocation sizes, so make powers of 2 generate the minimum required size.
const unsigned int HeapMaster::bucketSizes[] = {		// different bucket sizes
	16 + sizeof(Heap::Storage), 32 + sizeof(Heap::Storage), 48 + sizeof(Heap::Storage), 64 + sizeof(Heap::Storage), // 4
	96 + sizeof(Heap::Storage), 112 + sizeof(Heap::Storage), 128 + sizeof(Heap::Storage), // 3
	160, 192, 224, 256 + sizeof(Heap::Storage), // 4
	320, 384, 448, 512 + sizeof(Heap::Storage), // 4
	640, 768, 896, 1'024 + sizeof(Heap::Storage), // 4
	1'536, 2'048 + sizeof(Heap::Storage), // 2
	2'560, 3'072, 3'584, 4'096 + sizeof(Heap::Storage), // 4
	6'144, 8'192 + sizeof(Heap::Storage), // 2
	9'216, 10'240, 11'264, 12'288, 13'312, 14'336, 15'360, 16'384 + sizeof(Heap::Storage), // 8
	18'432, 20'480, 22'528, 24'576, 26'624, 28'672, 30'720, 32'768 + sizeof(Heap::Storage), // 8
	36'864, 40'960, 45'056, 49'152, 53'248, 57'344, 61'440, 65'536 + sizeof(Heap::Storage), // 8
	73'728, 81'920, 90'112, 98'304, 106'496, 114'688, 122'880, 131'072 + sizeof(Heap::Storage), // 8
	147'456, 163'840, 180'224, 196'608, 212'992, 229'376, 245'760, 262'144 + sizeof(Heap::Storage), // 8
	294'912, 327'680, 360'448, 393'216, 425'984, 458'752, 491'520, 524'288 + sizeof(Heap::Storage), // 8
	655'360, 786'432, 917'504, 1'048'576 + sizeof(Heap::Storage), // 4
	1'179'648, 1'310'720, 1'441'792, 1'572'864, 1'703'936, 1'835'008, 1'966'080, 2'097'152 + sizeof(Heap::Storage), // 8
	2'621'440, 3'145'728, 3'670'016, 4'194'304 + sizeof(Heap::Storage), // 4
};

static_assert( Heap::NoBucketSizes == sizeof(HeapMaster::bucketSizes) / sizeof(HeapMaster::bucketSizes[0]), "size of bucket array wrong" );


//####################### Memory Allocation Routines' Helpers ####################


#ifdef __STATISTICS__
static inline HeapStatistics & collectStats( HeapStatistics & stats ) {
	spin_acquire( &heapMaster.masterMgrLock );

	stats += heapMaster.stats;
	for ( Heap * node = heapMaster.heapManagersList; node; node = node->nextHeapManager ) {
		stats += node->stats;
	} // for

	spin_release(&heapMaster.masterMgrLock);
	return stats;
} // collectStats

// Use "write" because streams may be shutdown when calls are made.
static void printStats( HeapStatistics & stats ) {
	char helpText[1024];
	int len = snprintf( helpText, sizeof(helpText),
						"\nHeap statistics: (storage request / allocation)\n"
						"  malloc    >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
						"  aalloc    >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
						"  calloc    >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
						"  memalign  >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
						"  amemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
						"  cmemalign >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
						"  resize    >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
						"  realloc   >0 calls %'u; 0 calls %'u; storage %'llu / %'llu bytes\n"
						"  free      !null calls %'u; null calls %'u; storage %'llu / %'llu bytes\n"
						"  away      pulls %'u; pushes %'u; storage %'llu / %'llu bytes\n"
						"  sbrk      calls %'u; storage %'llu bytes\n"
						"  mmap      calls %'u; storage %'llu / %'llu bytes\n"
						"  munmap    calls %'u; storage %'llu / %'llu bytes\n"
						"  threads   started %'lu; exited %'lu\n"
						"  heaps     new %'lu; reused %'lu\n",
						stats.malloc_calls, stats.malloc_0_calls, stats.malloc_storage_request, stats.malloc_storage_alloc,
						stats.aalloc_calls, stats.aalloc_0_calls, stats.aalloc_storage_request, stats.aalloc_storage_alloc,
						stats.calloc_calls, stats.calloc_0_calls, stats.calloc_storage_request, stats.calloc_storage_alloc,
						stats.memalign_calls, stats.memalign_0_calls, stats.memalign_storage_request, stats.memalign_storage_alloc,
						stats.amemalign_calls, stats.amemalign_0_calls, stats.amemalign_storage_request, stats.amemalign_storage_alloc,
						stats.cmemalign_calls, stats.cmemalign_0_calls, stats.cmemalign_storage_request, stats.cmemalign_storage_alloc,
						stats.resize_calls, stats.resize_0_calls, stats.resize_storage_request, stats.resize_storage_alloc,
						stats.realloc_calls, stats.realloc_0_calls, stats.realloc_storage_request, stats.realloc_storage_alloc,
						stats.free_calls, stats.free_null_calls, stats.free_storage_request, stats.free_storage_alloc,
						stats.away_pulls, stats.away_pushes, stats.away_storage_request, stats.away_storage_alloc,
						heapMaster.sbrk_calls, heapMaster.sbrk_storage,
						stats.mmap_calls, stats.mmap_storage_request, stats.mmap_storage_alloc,
						stats.munmap_calls, stats.munmap_storage_request, stats.munmap_storage_alloc,
						heapMaster.threads_started, heapMaster.threads_exited,
						heapMaster.new_heap, heapMaster.reused_heap
		);
	NOWARNING( write( heapMaster.stats_fd, helpText, len ), unused-result );
} // printStats


static int printStatsXML( HeapStatistics & stats, FILE * stream ) {
	char helpText[1024];
	int len = snprintf( helpText, sizeof(helpText),
						"<malloc version=\"1\">\n"
						"<heap nr=\"0\">\n"
						"<sizes>\n"
						"</sizes>\n"
						"<total type=\"malloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
						"<total type=\"aalloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
						"<total type=\"calloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
						"<total type=\"memalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
						"<total type=\"amemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
						"<total type=\"cmemalign\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
						"<total type=\"resize\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
						"<total type=\"realloc\" >0 count=\"%'u;\" 0 count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
						"<total type=\"free\" !null=\"%'u;\" 0 null=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
						"<total type=\"away\" pulls=\"%'u;\" 0 pushes=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
						"<total type=\"sbrk\" count=\"%'u;\" size=\"%'llu\"/> bytes\n"
						"<total type=\"mmap\" count=\"%'u;\" size=\"%'llu / %'llu\" / > bytes\n"
						"<total type=\"munmap\" count=\"%'u;\" size=\"%'llu / %'llu\"/> bytes\n"
						"<total type=\"threads\" started=\"%'lu;\" exited=\"%'lu\"/>\n"
						"</malloc>",
						stats.malloc_calls, stats.malloc_0_calls, stats.malloc_storage_request, stats.malloc_storage_alloc,
						stats.aalloc_calls, stats.aalloc_0_calls, stats.aalloc_storage_request, stats.aalloc_storage_alloc,
						stats.calloc_calls, stats.calloc_0_calls, stats.calloc_storage_request, stats.calloc_storage_alloc,
						stats.memalign_calls, stats.memalign_0_calls, stats.memalign_storage_request, stats.memalign_storage_alloc,
						stats.amemalign_calls, stats.amemalign_0_calls, stats.amemalign_storage_request, stats.amemalign_storage_alloc,
						stats.cmemalign_calls, stats.cmemalign_0_calls, stats.cmemalign_storage_request, stats.cmemalign_storage_alloc,
						stats.resize_calls, stats.resize_0_calls, stats.resize_storage_request, stats.resize_storage_alloc,
						stats.realloc_calls, stats.realloc_0_calls, stats.realloc_storage_request, stats.realloc_storage_alloc,
						stats.free_calls, stats.free_null_calls, stats.free_storage_request, stats.free_storage_alloc,
						stats.away_pulls, stats.away_pushes, stats.away_storage_request, stats.away_storage_alloc,
						heapMaster.sbrk_calls, heapMaster.sbrk_storage,
						stats.mmap_calls, stats.mmap_storage_request, stats.mmap_storage_alloc,
						stats.munmap_calls, stats.munmap_storage_request, stats.munmap_storage_alloc,
						heapMaster.threads_started, heapMaster.threads_exited
		);
	NOWARNING( write( fileno(stream), helpText, len ), unused-result );
	return len;
} // printStatsXML
#endif // __STATISTICS__


inline void noMemory() {
	abort( "Heap memory exhausted at %zu bytes.\n"
		   "Possible cause is very large memory allocation and/or large amount of unfreed storage allocated by the program or system/library routines.",
		   ((char *)(sbrk( 0 )) - (char *)(heapMaster.heapBegin)) );
} // noMemory


static bool setMmapStart( size_t value ) {
  if ( value < __cfa_page_size || heapMaster.bucketSizes[Heap::NoBucketSizes - 1] < value ) return false;
	heapMaster.mmapStart = value;						// set global

	// find the closest bucket size less than or equal to the mmapStart size
	heapMaster.maxBucketsUsed = std::lower_bound( heapMaster.bucketSizes, heapMaster.bucketSizes + (Heap::NoBucketSizes - 1), heapMaster.mmapStart ) - heapMaster.bucketSizes; // binary search
	assert( heapMaster.maxBucketsUsed < Heap::NoBucketSizes ); // subscript failure ?
	assert( heapMaster.mmapStart <= heapMaster.bucketSizes[heapMaster.maxBucketsUsed] ); // search failure ?
	return true;
} // setMmapStart

// <-------+----------------------------------------------------> bsize (bucket size)
// |header |addr
//==================================================================================
//                   align/offset |
// <-----------------<------------+-----------------------------> bsize (bucket size)
//                   |fake-header | addr
#define headerAddr( addr ) ((Heap::Storage::Header *)( (char *)addr - sizeof(Heap::Storage) ))
#define realHeader( header ) ((Heap::Storage::Header *)((char *)header - header->kind.fake.offset))

// <-------<<--------------------- dsize ---------------------->> bsize (bucket size)
// |header |addr
//==================================================================================
//                   align/offset |
// <------------------------------<<---------- dsize --------->>> bsize (bucket size)
//                   |fake-header |addr
#define dataStorage( bsize, addr, header ) (bsize - ( (char *)addr - (char *)header ))


static inline void checkAlign( size_t alignment ) {
	if ( UNLIKELY( alignment < __ALIGN__ || ! Pow2( alignment ) ) ) {
		abort( "Alignment %zu for memory allocation is less than %d and/or not a power of 2.", alignment, __ALIGN__ );
	} // if
} // checkAlign


static inline void checkHeader( bool check, const char name[], void * addr ) {
	if ( UNLIKELY( check ) ) {							// bad address ?
		abort( "Attempt to %s storage %p with address outside the heap.\n"
			   "Possible cause is duplicate free on same block or overwriting of memory.",
			   name, addr );
	} // if
} // checkHeader


static inline void fakeHeader( Heap::Storage::Header *& header, size_t & alignment ) {
	if ( UNLIKELY( (header->kind.fake.alignment & 1) == 1 ) ) { // fake header ?
		alignment = header->kind.fake.alignment & -2;	// remove flag from value
		#ifdef __DEBUG__
		checkAlign( alignment );						// check alignment
		#endif // __DEBUG__
		header = realHeader( header );					// backup from fake to real header
	} else {
		alignment = __ALIGN__;							// => no fake header
	} // if
} // fakeHeader


static inline bool headers( const char name[] __attribute__(( unused )), void * addr, Heap::Storage::Header *& header, Heap::FreeHeader *& freeHead, size_t & size, size_t & alignment ) {
	header = headerAddr( addr );

  if ( UNLIKELY( addr < heapMaster.heapBegin || heapMaster.heapEnd < addr ) ) { // mmapped ?
		fakeHeader( header, alignment );
		size = header->kind.real.blockSize & -3;		// mmap size
		return true;
	} // if

	#ifdef __DEBUG__
	checkHeader( header < heapMaster.heapBegin, name, addr ); // bad low address ?
	#endif // __DEBUG__

	// header may be safe to dereference
	fakeHeader( header, alignment );
	#ifdef __DEBUG__
	checkHeader( header < heapMaster.heapBegin || heapMaster.heapEnd < header, name, addr ); // bad address ? (offset could be + or -)
	#endif // __DEBUG__

	freeHead = (Heap::FreeHeader *)((size_t)header->kind.real.home & -3);
	#ifdef __DEBUG__
	Heap * homeManager = freeHead->homeManager;
	if ( UNLIKELY( freeHead < &homeManager->freeLists[0] || &homeManager->freeLists[Heap::NoBucketSizes - 1] < freeHead ) ) {
		abort( "Attempt to %s storage %p with corrupted header.\n"
			   "Possible cause is duplicate free on same block or overwriting of header information.",
			   name, addr );
	} // if
	#endif // __DEBUG__
	size = freeHead->blockSize;
	return false;
} // headers


static inline void * master_extend( size_t size ) {
	spin_acquire( &heapMaster.masterExtLock );

	ptrdiff_t rem = heapMaster.heapRemaining - size;
	if ( UNLIKELY( rem < 0 ) ) {
		// If the size requested is bigger than the current remaining storage, increase the size of the heap.

		size_t increase = Ceiling( size > heapMaster.heapExpand ? size : heapMaster.heapExpand, __ALIGN__ );
		if ( UNLIKELY( sbrk( increase ) == (void *)-1 ) ) {	// failed, no memory ?
			spin_release( &heapMaster.masterExtLock );
			abort( NO_MEMORY_MSG, size );				// give up
		} // if
		#ifdef __STATISTICS__
		heapMaster.sbrk_calls += 1;
		heapMaster.sbrk_storage += increase;
		#endif // __STATISTICS__
		rem = heapMaster.heapRemaining + increase - size;
	} // if

	Heap::Storage * block = (Heap::Storage *)heapMaster.heapEnd;
	heapMaster.heapRemaining = rem;
	heapMaster.heapEnd = (char *)heapMaster.heapEnd + size;

	spin_release( &heapMaster.masterExtLock );
	return block;
} // master_extend


static inline void * manager_extend( size_t size ) {
	ptrdiff_t rem = heapManager->heapReserve - size;

	if ( UNLIKELY( rem < 0 ) ) {						// negative
		// If the size requested is bigger than the current remaining reserve, use the current reserve to populate
		// smaller freeLists, and increase the reserve.

		rem = heapManager->heapReserve;					// positive

		if ( rem >= heapMaster.bucketSizes[0] ) {
			Heap::FreeHeader * freeHead =
			#ifdef FASTLOOKUP
				rem < Heap::LookupSizes ? &(heapManager->freeLists[heapMaster.lookup[rem]]) :
			#endif // FASTLOOKUP
			std::lower_bound( heapManager->freeLists, heapManager->freeLists + heapMaster.maxBucketsUsed, rem ); // binary search

			if ( UNLIKELY( freeHead->blockSize > (size_t)rem ) ) freeHead -= 1;
			Heap::Storage * block = (Heap::Storage *)heapManager->heapBuffer;

			block->header.kind.real.next = freeHead->freeList;	// push on stack
			freeHead->freeList = block;
		} // if

		size_t increase = Ceiling( size > ( heapMaster.heapExpand / 10 ) ? size : ( heapMaster.heapExpand / 10 ), __ALIGN__ );
		heapManager->heapBuffer = master_extend(increase);
		rem = increase - size;
	} // if

	Heap::Storage * block = (Heap::Storage *)heapManager->heapBuffer;
	heapManager->heapReserve = rem;
	heapManager->heapBuffer = (char *)heapManager->heapBuffer + size;

	return block;
} // manager_extend


static inline void * doMalloc( size_t size
							  #ifdef __STATISTICS__
							  , unsigned int counter
							  #endif // __STATISTICS__
		) {
	 Heap::Storage * block;

	// Look up size in the size list.  Make sure the user request includes space for the header that must be allocated
	// along with the block and is a multiple of the alignment size.
	size_t tsize = size + sizeof(Heap::Storage);

	#ifdef __STATISTICS__
	heapManager->stats.counters[counter].cnt1 += 1;
	heapManager->stats.counters[counter].cnt3 += size;
	#endif // __STATISTICS__

	if ( LIKELY( tsize < heapMaster.mmapStart ) ) {		// small size => sbrk
		Heap::FreeHeader * freeHead =
			#ifdef FASTLOOKUP
			LIKELY( tsize < Heap::LookupSizes ) ? &(heapManager->freeLists[heapMaster.lookup[tsize]]) :
			#endif // FASTLOOKUP
			std::lower_bound( heapManager->freeLists, heapManager->freeLists + heapMaster.maxBucketsUsed, tsize ); // binary search

		assert( freeHead <= &heapManager->freeLists[heapMaster.maxBucketsUsed] ); // subscripting error ?
		assert( tsize <= freeHead->blockSize );			// search failure ?
		tsize = freeHead->blockSize;					// total space needed for request
		#ifdef __STATISTICS__
		heapManager->stats.counters[counter].cnt4 += tsize;
		#endif // __STATISTICS__

		block = freeHead->freeList;						// remove node from stack
		if ( UNLIKELY( block == nullptr ) ) {			// no free block ?
			// Freelist for that size is empty, so carve it out of the heap, if there is enough left, or get some more
			// and then carve it off.

			#ifdef AWAYSPIN
			spin_acquire( &freeHead->awayLock );
			block = freeHead->awayList;
			freeHead->awayList = nullptr;
			spin_release( &freeHead->awayLock );
			#else
			block = __atomic_exchange_n( &freeHead->awayList, nullptr, __ATOMIC_SEQ_CST );
			#endif // AWAYSPIN
			if ( LIKELY( block == nullptr ) ) {			// away list also empty?
				block = (Heap::Storage *)manager_extend( tsize ); // mutual exclusion on call
			} else {									// merge awayList into freeHead
				#ifdef __STATISTICS__
				heapManager->stats.away_pulls += 1;
				#endif // __STATISTICS__
				freeHead->freeList = block->header.kind.real.next;
			} // if
		} else {
			freeHead->freeList = block->header.kind.real.next;
		} // if

		block->header.kind.real.home = freeHead;		// pointer back to free list of apropriate size
	} else {											// large size => mmap
  if ( UNLIKELY( size > ULONG_MAX - __cfa_page_size ) ) return nullptr; // error check
		tsize = Ceiling( tsize, __cfa_page_size );	// must be multiple of page size
		#ifdef __STATISTICS__
		heapManager->stats.counters[counter].cnt4 += tsize;
		heapManager->stats.mmap_calls += 1;
		heapManager->stats.mmap_storage_request += size;
		heapManager->stats.mmap_storage_alloc += tsize;
		#endif // __STATISTICS__

		block = (Heap::Storage *)::mmap( 0, tsize, __map_prot, MAP_PRIVATE | MAP_ANONYMOUS, heapMaster.mmapFd, 0 );
		if ( UNLIKELY( block == MAP_FAILED ) ) {		// failed ?
			if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, tsize ); // no memory
			// Do not call strerror( errno ) as it may call malloc.
			abort( "(Heap &)0x%p.doMalloc() : internal error, mmap failure, size:%zu %lu %lu error %d.",
				   &heapManager, tsize, size, heapMaster.mmapStart, errno );
		} // if
		block->header.kind.real.blockSize = tsize;		// storage size for munmap
	} // if

	block->header.kind.real.size = size;				// store allocation size
	void * addr = &(block->data);						// adjust off header to user bytes
	assert( ((uintptr_t)addr & (__ALIGN__ - 1)) == 0 ); // minimum alignment ?

	#ifdef __DEBUG__
	heapManager->allocUnfreed += size;
	#endif // __DEBUG__

	return addr;
} // doMalloc


static inline void doFree( void * addr ) {
	DISABLE_INTERRUPTS;

	if ( UNLIKELY( ! heapManagerBootFlag ) ) Heap::heapManagerCtor( // trigger for first heap
		#ifdef __DEBUG__
		0
		#endif // __DEBUG__
		);

	Heap::Storage::Header * header;
	Heap::FreeHeader * freeHead;
	size_t size, alignment;								// not used (see realloc)

	bool mapped = headers( "free", addr, header, freeHead, size, alignment );

	if ( UNLIKELY( mapped ) ) {						// mmapped ?
		#ifdef __STATISTICS__
		heapManager->stats.munmap_calls += 1;
		heapManager->stats.munmap_storage_request += header->kind.real.size;
		heapManager->stats.munmap_storage_alloc += size;
		#endif // __STATISTICS__
		if ( UNLIKELY( munmap( header, size ) == -1 ) ) {
			abort( "Attempt to deallocate storage %p not allocated or with corrupt header.\n"
				   "Possible cause is invalid pointer.",
				   addr );
		} // if
	} else {
		if ( LIKELY( heapManager == freeHead->homeManager ) ) {	// belongs to this thread
			header->kind.real.next = freeHead->freeList; // push on stack
			freeHead->freeList = (Heap::Storage *)header;
		} else {										// return to thread owner
			#ifdef AWAYSPIN
			spin_acquire( &freeHead->awayLock );
			header->kind.real.next = freeHead->awayList; // push to bucket away list
			freeHead->awayList = (Heap::Storage *)header;
			spin_release( &freeHead->awayLock );
			#else										// lock free
			header->kind.real.next = freeHead->awayList; // link new node to top node
			// CAS resets header->kind.real.next = freeHead->awayList on failure
			while ( ! __atomic_compare_exchange_n( &freeHead->awayList, &header->kind.real.next, header,
												   false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST ) );
			#endif // AWAYSPIN

			// detect free after thread-local storage destruction and use global stats in that case
			if ( UNLIKELY( heapManager == nullptr ) ) {
				#ifdef __STATISTICS__
				AtomicFetchAdd( heapMaster.stats.free_storage_request, header->kind.real.size );
				AtomicFetchAdd( heapMaster.stats.free_storage_alloc, size );
				#endif // __STATISTICS__
				// away push counters are not incremented because this is a self-away push, and there is no
				// corresponding pull counter that needs to match.
				ENABLE_INTERRUPTS;
				return;
			} // if

			#ifdef __STATISTICS__
			heapManager->stats.away_pushes += 1;
			heapManager->stats.away_storage_request += header->kind.real.size;
			heapManager->stats.away_storage_alloc += size;
			#endif // __STATISTICS__
		} // if
	} // if

	#ifdef __STATISTICS__
	heapManager->stats.free_storage_request += header->kind.real.size;
	heapManager->stats.free_storage_alloc += size;
	#endif // __STATISTICS__

	#ifdef __DEBUG__
	heapManager->allocUnfreed -= header->kind.real.size;
	#endif // __DEBUG__

	ENABLE_INTERRUPTS;
} // doFree


static inline void * mallocNoStats( size_t size
									#ifdef __STATISTICS__
									, unsigned int counter
									#endif // __STATISTICS__
		) {
	DISABLE_INTERRUPTS;

	if ( UNLIKELY( ! heapManagerBootFlag ) ) Heap::heapManagerCtor( // trigger for first heap
		#ifdef __DEBUG__
		size
		#endif // __DEBUG__
		);

  if ( UNLIKELY( size ) == 0 ||							// 0 BYTE ALLOCATION RETURNS NULL POINTER
	   UNLIKELY( size > ULONG_MAX - sizeof(Heap::Storage) ) ) { // error check
		#ifdef __STATISTICS__
		heapManager->stats.counters[counter].cnt2 += 1;
		#endif // __STATISTICS__
		ENABLE_INTERRUPTS;
		return nullptr;
	} // if

	void * addr = doMalloc( size
							#ifdef __STATISTICS__
							, counter
							#endif // __STATISTICS__
		);

	ENABLE_INTERRUPTS;
	return addr;
} // mallocNoStats


static inline void * memalignNoStats( size_t alignment, size_t size
									  #ifdef __STATISTICS__
									  , unsigned int counter
									  #endif // __STATISTICS__
		) {
	DISABLE_INTERRUPTS;

	if ( UNLIKELY( ! heapManagerBootFlag ) ) Heap::heapManagerCtor( // trigger for first heap
		#ifdef __DEBUG__
		size
		#endif // __DEBUG__
		);

  if ( UNLIKELY( size ) == 0 ||							// 0 BYTE ALLOCATION RETURNS NULL POINTER
	   UNLIKELY( size > ULONG_MAX - sizeof(Heap::Storage) ) ) { // error check
		#ifdef __STATISTICS__
		heapManager->stats.counters[counter].cnt2 += 1;
		#endif // __STATISTICS__

		ENABLE_INTERRUPTS;
		return nullptr;
	} // if

	#ifdef __DEBUG__
	checkAlign( alignment );							// check alignment
	#endif // __DEBUG__

	// if alignment <= default alignment, do normal malloc as two headers are unnecessary
  if ( UNLIKELY( alignment <= __ALIGN__ ) ) {
		void * addr = doMalloc( size
			#ifdef __STATISTICS__
			, counter
			#endif // __STATISTICS__
		);

		ENABLE_INTERRUPTS;
		return addr;
  }


	// Allocate enough storage to guarantee an address on the alignment boundary, and sufficient space before it for
	// administrative storage. NOTE, WHILE THERE ARE 2 HEADERS, THE FIRST ONE IS IMPLICITLY CREATED BY DOMALLOC.
	//      .-------------v-----------------v----------------v----------,
	//      | Real Header | ... padding ... |   Fake Header  | data ... |
	//      `-------------^-----------------^-+--------------^----------'
	//      |<--------------------------------' offset/align |<-- alignment boundary

	// subtract __ALIGN__ because it is already the minimum alignment
	// add sizeof(Heap::Storage) for fake header
	char * addr = (char *)doMalloc( size + alignment - __ALIGN__ + sizeof(Heap::Storage)
									#ifdef __STATISTICS__
									, counter
									#endif // __STATISTICS__
		);

	// address in the block of the "next" alignment address
	char * user = (char *)Ceiling( (uintptr_t)(addr + sizeof(Heap::Storage)), alignment );

	// address of header from malloc
	Heap::Storage::Header * realHeader = headerAddr( addr );
	realHeader->kind.real.size = size;					// correct size to eliminate above alignment offset
	// address of fake header * before* the alignment location
	Heap::Storage::Header * fakeHeader = headerAddr( user );
	// SKULLDUGGERY: insert the offset to the start of the actual storage block and remember alignment
	fakeHeader->kind.fake.offset = (char *)fakeHeader - (char *)realHeader;
	// SKULLDUGGERY: odd alignment imples fake header
	fakeHeader->kind.fake.alignment = alignment | 1;

	ENABLE_INTERRUPTS;
	return user;
} // memalignNoStats

// Operators new and new [] call malloc; delete calls free


//####################### Memory Allocation Routines ####################


extern "C" {
	// Allocates size bytes and returns a pointer to the allocated memory.  The contents are undefined. If size is 0,
	// then malloc() returns a unique pointer value that can later be successfully passed to free().
	void * malloc( size_t size ) {
		return mallocNoStats( size
							  #ifdef __STATISTICS__
							  , HeapStatistics::MALLOC
							  #endif // __STATISTICS__
			);
	} // malloc


	// Same as malloc() except size bytes is an array of dim elements each of elemSize bytes.
	void * aalloc( size_t dim, size_t elemSize ) {
		return mallocNoStats( dim * elemSize
							  #ifdef __STATISTICS__
							  , HeapStatistics::AALLOC
							  #endif // __STATISTICS__
			);
	} // aalloc


	// Same as aalloc() with memory set to zero.
	void * calloc( size_t dim, size_t elemSize ) {
		size_t size = dim * elemSize;
		char * addr = (char *)mallocNoStats( size
											 #ifdef __STATISTICS__
											 , HeapStatistics::CALLOC
											 #endif // __STATISTICS__
			);

		if ( UNLIKELY( addr == NULL ) ) return NULL; // stop further processing if 0p is returned

		Heap::Storage::Header * header;
		Heap::FreeHeader * freeHead;
		size_t bsize, alignment;

		#ifndef __DEBUG__
		bool mapped =
		#endif // __DEBUG__
			headers( "calloc", addr, header, freeHead, bsize, alignment );

		#ifndef __DEBUG__
		// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
		if ( LIKELY( ! mapped ) )
		#endif // __DEBUG__
			// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
			// `-header`-addr                      `-size
			memset( addr, '\0', size );					// set to zeros

		header->kind.real.blockSize |= 2;				// mark as zero filled
		return addr;
	} // calloc


	// Change the size of the memory block pointed to by oaddr to size bytes. The contents are undefined.  If oaddr is
	// nullptr, then the call is equivalent to malloc(size), for all values of size; if size is equal to zero, and oaddr is
	// not nullptr, then the call is equivalent to free(oaddr). Unless oaddr is nullptr, it must have been returned by an earlier
	// call to malloc(), alloc(), calloc() or realloc(). If the area pointed to was moved, a free(oaddr) is done.
	void * resize( void * oaddr, size_t size ) {
	  if ( UNLIKELY( oaddr == nullptr ) ) {				// special cases
			return mallocNoStats( size
								  #ifdef __STATISTICS__
								  , HeapStatistics::RESIZE
								  #endif // __STATISTICS__
				);
		} // if

		// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
	  if ( UNLIKELY( size == 0 ) ) {					// special cases
			#ifdef __STATISTICS__
			heapManager->stats.resize_0_calls += 1;
			#endif // __STATISTICS__
			doFree( oaddr );							// free previous storage
			return nullptr;
		} // if

		Heap::Storage::Header * header;
		Heap::FreeHeader * freeHead;
		size_t bsize, oalign;
		headers( "resize", oaddr, header, freeHead, bsize, oalign );

		size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
		// same size, DO NOT preserve STICKY PROPERTIES.
		if ( oalign == __ALIGN__ && size <= odsize && odsize <= size * 2 ) { // allow 50% wasted storage for smaller size
			#ifdef __STATISTICS__
			heapManager->stats.resize_calls += 1;
			#endif // __STATISTICS__
			header->kind.real.blockSize &= -2;			// no alignment and turn off 0 fill
			header->kind.real.size = size;				// reset allocation size
			return oaddr;
		} // if

		// change size, DO NOT preserve STICKY PROPERTIES.
		doFree( oaddr );								// free previous storage
		return mallocNoStats( size						// create new area
							  #ifdef __STATISTICS__
							  , HeapStatistics::RESIZE
							  #endif // __STATISTICS__
			);
	} // resize


	// Same as resize() but the contents are unchanged in the range from the start of the region up to the minimum of
	// the old and new sizes.
	void * realloc( void * oaddr, size_t size ) {
	  if ( UNLIKELY( oaddr == nullptr ) ) {				// special cases
			return mallocNoStats( size
								  #ifdef __STATISTICS__
								  , HeapStatistics::REALLOC
								  #endif // __STATISTICS__
				);
		} // if

		// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
	  if ( UNLIKELY( size == 0 ) ) {					// special cases
			#ifdef __STATISTICS__
			heapManager->stats.realloc_0_calls += 1;
			#endif // __STATISTICS__
			doFree( oaddr );							// free previous storage
			return nullptr;
		} // if

		Heap::Storage::Header * header;
		Heap::FreeHeader * freeHead;
		size_t bsize, oalign;
		headers( "realloc", oaddr, header, freeHead, bsize, oalign );

		size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket
		size_t osize = header->kind.real.size;			// old allocation size
		bool ozfill = (header->kind.real.blockSize & 2); // old allocation zero filled
	  if ( UNLIKELY( size <= odsize ) && odsize <= size * 2 ) { // allow up to 50% wasted storage
			#ifdef __STATISTICS__
			heapManager->stats.realloc_calls += 1;
			heapManager->stats.realloc_storage_request += size;
			#endif // __STATISTICS__

	  		header->kind.real.size = size;				// reset allocation size
	  		if ( UNLIKELY( ozfill ) && size > osize ) {	// previous request zero fill and larger ?
	  			memset( (char *)oaddr + osize, '\0', size - osize ); // initialize added storage
	  		} // if
			return oaddr;
		} // if

		// change size and copy old content to new storage

		void * naddr;
		if ( UNLIKELY( oalign <= __ALIGN__ ) ) {		// previous request not aligned ?
			naddr = mallocNoStats( size					// create new area
								   #ifdef __STATISTICS__
								   , HeapStatistics::REALLOC
								   #endif // __STATISTICS__
				);
		} else {
			naddr = memalignNoStats( oalign, size		// create new aligned area
									 #ifdef __STATISTICS__
									 , HeapStatistics::REALLOC
									 #endif // __STATISTICS__
				);
		} // if

		headers( "realloc", naddr, header, freeHead, bsize, oalign );
		// To preserve prior fill, the entire bucket must be copied versus the size.
		memcpy( naddr, oaddr, std::min( osize, size ) ); // copy bytes
		doFree( oaddr );								// free previous storage

		if ( UNLIKELY( ozfill ) ) {						// previous request zero fill ?
			header->kind.real.blockSize |= 2;			// mark new request as zero filled
			if ( size > osize ) {						// previous request larger ?
				memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
			} // if
		} // if
		return naddr;
	} // realloc


	// Same as malloc() except the memory address is a multiple of alignment, which must be a power of two. (obsolete)
	void * memalign( size_t alignment, size_t size ) {
		return memalignNoStats( alignment, size
								#ifdef __STATISTICS__
								, HeapStatistics::MEMALIGN
								#endif // __STATISTICS__
			);
	} // memalign


	// Same as aalloc() with memory alignment.
	void * amemalign( size_t alignment, size_t dim, size_t elemSize ) {
		return memalignNoStats( alignment, dim * elemSize
								#ifdef __STATISTICS__
								, HeapStatistics::AMEMALIGN
								#endif // __STATISTICS__
			);
	} // amemalign


	// Same as calloc() with memory alignment.
	void * cmemalign( size_t alignment, size_t dim, size_t elemSize ) {
		size_t size = dim * elemSize;
		char * addr = (char *)memalignNoStats( alignment, size
											   #ifdef __STATISTICS__
											   , HeapStatistics::CMEMALIGN
											   #endif // __STATISTICS__
			);

		if ( UNLIKELY( addr == NULL ) ) return NULL; // stop further processing if 0p is returned

		Heap::Storage::Header * header;
		Heap::FreeHeader * freeHead;
		size_t bsize;

		#ifndef __DEBUG__
		bool mapped =
		#endif // __DEBUG__
			headers( "cmemalign", addr, header, freeHead, bsize, alignment );

		// Mapped storage is zero filled, but in debug mode mapped memory is scrubbed in doMalloc, so it has to be reset to zero.
		#ifndef __DEBUG__
		if ( LIKELY( ! mapped ) )
		#endif // __DEBUG__
			// <-------0000000000000000000000000000UUUUUUUUUUUUUUUUUUUUUUUUU> bsize (bucket size) U => undefined
			// `-header`-addr                      `-size
			memset( addr, '\0', size );					// set to zeros

		header->kind.real.blockSize |= 2;				// mark as zero filled
		return addr;
	} // cmemalign


	// Same as memalign(), but ISO/IEC 2011 C11 Section 7.22.2 states: the value of size shall be an integral multiple
	// of alignment. This requirement is universally ignored.
	void * aligned_alloc( size_t alignment, size_t size ) {
		return memalign( alignment, size );
	} // aligned_alloc


	// Allocates size bytes and places the address of the allocated memory in *memptr. The address of the allocated
	// memory shall be a multiple of alignment, which must be a power of two and a multiple of sizeof(void *). If size
	// is 0, then posix_memalign() returns either nullptr, or a unique pointer value that can later be successfully passed to
	// free(3).
	int posix_memalign( void ** memptr, size_t alignment, size_t size ) {
	  if ( UNLIKELY( alignment < __ALIGN__ || ! Pow2( alignment ) ) ) return EINVAL; // check alignment
		*memptr = memalign( alignment, size );
		return 0;
	} // posix_memalign


	// Allocates size bytes and returns a pointer to the allocated memory. The memory address shall be a multiple of the
	// page size.  It is equivalent to memalign(sysconf(_SC_PAGESIZE),size).
	void * valloc( size_t size ) {
		return memalign( __cfa_page_size, size );
	} // valloc


	// Same as valloc but rounds size to multiple of page size.
	void * pvalloc( size_t size ) {						// round size to multiple of page size
		return memalign( __cfa_page_size, Ceiling( size, __cfa_page_size ) );
	} // pvalloc


	// Frees the memory space pointed to by ptr, which must have been returned by a previous call to malloc(), calloc()
	// or realloc().  Otherwise, or if free(ptr) has already been called before, undefined behaviour occurs. If ptr is
	// nullptr, no operation is performed.
	void free( void * addr ) {
		// detect free after thread-local storage destruction and use global stats in that case
		#ifdef __STATISTICS__
		DISABLE_INTERRUPTS;
		if ( UNLIKELY( ! heapManagerBootFlag ) ) Heap::heapManagerCtor( // trigger for first heap
			#ifdef __DEBUG__
			0
			#endif // __DEBUG__
			);
		#endif // __STATISTICS__

		// detect free after thread-local storage destruction and use global stats in that case
		if ( UNLIKELY( addr == nullptr ) ) {			// special case
			#ifdef __STATISTICS__
			if ( LIKELY( heapManager ) ) heapManager->stats.free_null_calls += 1;
			else AtomicFetchAdd( heapMaster.stats.free_null_calls, 1 );
			ENABLE_INTERRUPTS;
			#endif // __STATISTICS__
			return;
		} // fi

		#ifdef __STATISTICS__
		if ( LIKELY( heapManager ) ) heapManager->stats.free_calls += 1;
		else AtomicFetchAdd( heapMaster.stats.free_calls, 1 );
		ENABLE_INTERRUPTS;
		#endif // __STATISTICS__

		doFree( addr );
	} // free


	// Returns the alignment of an allocation.
	size_t malloc_alignment( void * addr ) {
	  if ( UNLIKELY( addr == nullptr ) ) return __ALIGN__; // minimum alignment
		Heap::Storage::Header * header = headerAddr( addr );
		if ( UNLIKELY( (header->kind.fake.alignment & 1) == 1 ) ) {	// fake header ?
			return header->kind.fake.alignment & -2;	// remove flag from value
		} else {
			return __ALIGN__;							// minimum alignment
		} // if
	} // malloc_alignment


	// Returns true if the allocation is zero filled, e.g., allocated by calloc().
	bool malloc_zero_fill( void * addr ) {
	  if ( UNLIKELY( addr == nullptr ) ) return false;	// null allocation is not zero fill
		Heap::Storage::Header * header = headerAddr( addr );
		if ( UNLIKELY( (header->kind.fake.alignment & 1) == 1 ) ) { // fake header ?
			header = realHeader( header );				// backup from fake to real header
		} // if
		return (header->kind.real.blockSize & 2) != 0;	// zero filled ?
	} // malloc_zero_fill


	// Returns original total allocation size (not bucket size) => array size is dimension * sizeof(T).
	size_t malloc_size( void * addr ) {
	  if ( UNLIKELY( addr == nullptr ) ) return 0;		// null allocation is not zero fill
		Heap::Storage::Header * header = headerAddr( addr );
		if ( UNLIKELY( (header->kind.fake.alignment & 1) == 1 ) ) { // fake header ?
			header = realHeader( header );				// backup from fake to real header
		} // if
		return header->kind.real.size;
	} // malloc_size


	// Returns the number of usable bytes in the block pointed to by ptr, a pointer to a block of memory allocated by
	// malloc or a related function.
	size_t malloc_usable_size( void * addr ) {
	  if ( UNLIKELY( addr == nullptr ) ) return 0;		// null allocation has 0 size
		Heap::Storage::Header * header;
		Heap::FreeHeader * freeHead;
		size_t bsize, alignment;

		headers( "malloc_usable_size", addr, header, freeHead, bsize, alignment );
		return dataStorage( bsize, addr, header );		// data storage in bucket
	} // malloc_usable_size


	// Prints (on default standard error) statistics about memory allocated by malloc and related functions.
	void malloc_stats() {
		#ifdef __STATISTICS__
		HeapStatistics stats;
		printStats( collectStats( stats ) );
		#else
		#define MALLOC_STATS_MSG "malloc_stats statistics disabled.\n"
		NOWARNING( write( STDERR_FILENO, MALLOC_STATS_MSG, sizeof( MALLOC_STATS_MSG ) - 1 /* size includes '\0' */ ), unused-result );
		#endif // __STATISTICS__
	} // malloc_stats


	// Changes the file descriptor where malloc_stats() writes statistics.
	int malloc_stats_fd( int fd __attribute__(( unused )) ) {
		#ifdef __STATISTICS__
		int temp = heapMaster.stats_fd;
		heapMaster.stats_fd = fd;
		return temp;
		#else
		return -1;										// unsupported
		#endif // __STATISTICS__
	} // malloc_stats_fd


	// Prints an XML string that describes the current state of the memory-allocation implementation in the caller.
	// The string is printed on the file stream stream.  The exported string includes information about all arenas (see
	// malloc).
	int malloc_info( int options, FILE * stream __attribute__(( unused )) ) {
	  if ( options != 0 ) { errno = EINVAL; return -1; }
		#ifdef __STATISTICS__
		HeapStatistics stats;
		return printStatsXML( collectStats( stats ), stream );
		#else
		return 0;										// unsupported
		#endif // __STATISTICS__
	} // malloc_info


	// Adjusts parameters that control the behaviour of the memory-allocation functions (see malloc). The param argument
	// specifies the parameter to be modified, and value specifies the new value for that parameter.
	int mallopt( int option, int value ) {
	  if ( value < 0 ) return 0;
		switch( option ) {
		  case M_TOP_PAD:
			heapMaster.heapExpand = Ceiling( value, __cfa_page_size );
			return 1;
		  case M_MMAP_THRESHOLD:
			if ( setMmapStart( value ) ) return 1;
			break;
		} // switch
		return 0;										// error, unsupported
	} // mallopt


	// Attempt to release free memory at the top of the heap (by calling sbrk with a suitable argument).
	int malloc_trim( size_t ) {
		return 0;										// => impossible to release memory
	} // malloc_trim


	// Records the current state of all malloc internal bookkeeping variables (but not the actual contents of the heap
	// or the state of malloc_hook functions pointers).  The state is recorded in a system-dependent opaque data
	// structure dynamically allocated via malloc, and a pointer to that data structure is returned as the function
	// result.  (The caller must free this memory.)
	void * malloc_get_state( void ) {
		return nullptr;									// unsupported
	} // malloc_get_state


	// Restores the state of all malloc internal bookkeeping variables to the values recorded in the opaque data
	// structure pointed to by state.
	int malloc_set_state( void * ) {
		return 0;										// unsupported
	} // malloc_set_state

	// Set the amount (bytes) to extend the heap size once all the current storage in the heap is allocated.
	size_t malloc_expansion() { return __DEFAULT_HEAP_EXPANSION__; }

	// Set the crossover point between allocations occuring in the sbrk area or separately mmapped.
	size_t malloc_mmap_start() { return __DEFAULT_MMAP_START__; }

	// Amount subtracted to adjust for unfreed program storage (debug only).
	size_t malloc_unfreed() { return __DEFAULT_HEAP_UNFREED__; }
} // extern "C"


#ifdef __cforall
void * resize( void * oaddr, size_t nalign, size_t size )
#else
extern "C" {
void * _X6resizeFPv_Pvmm__1( void * oaddr, size_t nalign, size_t size )
#endif
{
  if ( UNLIKELY( oaddr == nullptr ) ) {
		return memalignNoStats( nalign, size
								#ifdef __STATISTICS__
								, HeapStatistics::RESIZE
								#endif // __STATISTICS__
			);
	} // if

	// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
  if ( UNLIKELY( size == 0 ) ) {						// special cases
		#ifdef __STATISTICS__
		heapManager->stats.resize_0_calls += 1;
		#endif // __STATISTICS__
		doFree( oaddr );								// free previous storage
		return nullptr;
	} // if

	#ifdef __DEBUG__
	checkAlign( nalign );								// check alignment
	#endif // __DEBUG__

	// Attempt to reuse existing alignment.
	Heap::Storage::Header * header = headerAddr( oaddr );
	bool isFakeHeader = header->kind.fake.alignment & 1; // old fake header ?
	size_t oalign;
	if ( UNLIKELY( isFakeHeader ) ) {
		oalign = header->kind.fake.alignment & -2;		// old alignment
		if ( UNLIKELY( (uintptr_t)oaddr % nalign == 0	// lucky match ?
			 && ( oalign <= nalign						// going down
				  || (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
			) ) {
			headerAddr( oaddr )->kind.fake.alignment = nalign | 1; // update alignment (could be the same)
			Heap::FreeHeader * freeHead;
			size_t bsize, oalign;
			headers( "resize", oaddr, header, freeHead, bsize, oalign );
			size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket

			if ( size <= odsize && odsize <= size * 2 ) { // allow 50% wasted data storage
				headerAddr( oaddr )->kind.fake.alignment = nalign | 1; // update alignment (could be the same)

				header->kind.real.blockSize &= -2;		// turn off 0 fill
				header->kind.real.size = size;			// reset allocation size
				return oaddr;
			} // if
		} // if
	} else if ( ! isFakeHeader							// old real header (aligned on libAlign) ?
				&& nalign == __ALIGN__ ) {				// new alignment also on libAlign => no fake header needed
		return resize( oaddr, size );					// duplicate special case checks
	} // if

	// change size, DO NOT preserve STICKY PROPERTIES.
	doFree( oaddr );									// free previous storage
	return memalignNoStats( nalign, size				// create new aligned area
							#ifdef __STATISTICS__
							, HeapStatistics::RESIZE
							#endif // __STATISTICS__
		);
} // resize
#ifndef __cforall
}
#endif


#ifdef __cforall
void * realloc( void * oaddr, size_t nalign, size_t size )
#else
extern "C" {
void * _X7reallocFPv_Pvmm__1( void * oaddr, size_t nalign, size_t size )
#endif
{
  if ( UNLIKELY( oaddr == nullptr ) ) {
		return memalignNoStats( nalign, size
								#ifdef __STATISTICS__
								, HeapStatistics::REALLOC
								#endif // __STATISTICS__
			);
	} // if

	// If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
  if ( UNLIKELY( size == 0 ) ) {						// special cases
		#ifdef __STATISTICS__
		heapManager->stats.realloc_0_calls += 1;
		#endif // __STATISTICS__
		doFree( oaddr );								// free previous storage
		return nullptr;
	} // if

	#ifdef __DEBUG__
	checkAlign( nalign );								// check alignment
	#endif // __DEBUG__

	// Attempt to reuse existing alignment.
	Heap::Storage::Header * header = headerAddr( oaddr );
	bool isFakeHeader = header->kind.fake.alignment & 1; // old fake header ?
	size_t oalign;
	if ( UNLIKELY( isFakeHeader ) ) {
		oalign = header->kind.fake.alignment & -2;		// old alignment
		if ( UNLIKELY( (uintptr_t)oaddr % nalign == 0	// lucky match ?
			 && ( oalign <= nalign						// going down
				  || (oalign >= nalign && oalign <= 256) ) // little alignment storage wasted ?
			) ) {
			headerAddr( oaddr )->kind.fake.alignment = nalign | 1; // update alignment (could be the same)
			return realloc( oaddr, size );				// duplicate special case checks
		} // if
	} else if ( ! isFakeHeader							// old real header (aligned on libAlign) ?
				&& nalign == __ALIGN__ ) {				// new alignment also on libAlign => no fake header needed
		return realloc( oaddr, size );					// duplicate special case checks
	} // if

	Heap::FreeHeader * freeHead;
	size_t bsize;
	headers( "realloc", oaddr, header, freeHead, bsize, oalign );

	// change size and copy old content to new storage

	size_t osize = header->kind.real.size;				// old allocation size
	bool ozfill = (header->kind.real.blockSize & 2);	// old allocation zero filled

	void * naddr = memalignNoStats( nalign, size		// create new aligned area
									#ifdef __STATISTICS__
									, HeapStatistics::REALLOC
									#endif // __STATISTICS__
		);

	headers( "realloc", naddr, header, freeHead, bsize, oalign );
	memcpy( naddr, oaddr, std::min( osize, size ) );	// copy bytes
	doFree( oaddr );									// free previous storage

	if ( UNLIKELY( ozfill ) ) {							// previous request zero fill ?
		header->kind.real.blockSize |= 2;				// mark new request as zero filled
		if ( size > osize ) {							// previous request larger ?
			memset( (char *)naddr + osize, '\0', size - osize ); // initialize added storage
		} // if
	} // if
	return naddr;
} // realloc
#ifndef __cforall
}
#endif

// zip -r HeapPerThread.zip heap/HeapPerThread.h heap/HeapPerThread.cc heap/Makefile heap/affinity.h heap/test.cc heap/away.cc

// g++-10 -Wall -Wextra -g -O3 -DNDEBUG -D__STATISTICS__ -DTLS HeapPerThread.cc -fPIC -shared -o HeapPerThread.so

// Local Variables: //
// tab-width: 4 //
// compile-command: "g++-10 -Wall -Wextra -g -O3 -DNDEBUG -D__STATISTICS__ HeapPerThread.cc -c" //
// End: //