source: libcfa/src/heap.cfa @ b4aa1ab

//
// Cforall Version 1.0.0 Copyright (C) 2017 University of Waterloo
//
// The contents of this file are covered under the licence agreement in the
// file "LICENCE" distributed with Cforall.
//
// heap.cfa --
//
// Author           : Peter A. Buhr
// Created On       : Tue Dec 19 21:58:35 2017
// Last Modified By : Peter A. Buhr
// Last Modified On : Sun Dec 13 22:04:10 2020
// Update Count     : 984
//

#include <unistd.h>                                                                             // sbrk, sysconf
#include <stdbool.h>                                                                    // true, false
#include <stdio.h>                                                                              // snprintf, fileno
#include <errno.h>                                                                              // errno
#include <string.h>                                                                             // memset, memcpy
#include <limits.h>                                                                             // ULONG_MAX
#include <malloc.h>                                                                             // memalign, malloc_usable_size
#include <sys/mman.h>                                                                   // mmap, munmap

#include "bits/align.hfa"                                                               // libAlign
#include "bits/defs.hfa"                                                                // likely, unlikely
#include "bits/locks.hfa"                                                               // __spinlock_t
#include "startup.hfa"                                                                  // STARTUP_PRIORITY_MEMORY
#include "math.hfa"                                                                             // ceiling
#include "bitmanip.hfa"                                                                 // is_pow2, ceiling2

static bool traceHeap = false;

inline bool traceHeap() { return traceHeap; }

bool traceHeapOn() {
        bool temp = traceHeap;
        traceHeap = true;
        return temp;
} // traceHeapOn

bool traceHeapOff() {
        bool temp = traceHeap;
        traceHeap = false;
        return temp;
} // traceHeapOff

bool traceHeapTerm() { return false; }


static bool prtFree = false;

inline bool prtFree() {
        return prtFree;
} // prtFree

bool prtFreeOn() {
        bool temp = prtFree;
        prtFree = true;
        return temp;
} // prtFreeOn

bool prtFreeOff() {
        bool temp = prtFree;
        prtFree = false;
        return temp;
} // prtFreeOff


enum {
        // Define the default extension heap amount in units of bytes. When the uC++ supplied heap reaches the brk address,
        // the brk address is extended by the extension amount.
        __CFA_DEFAULT_HEAP_EXPANSION__ = (1 * 1024 * 1024),

        // Define the mmap crossover point during allocation. Allocations less than this amount are allocated from buckets;
        // values greater than or equal to this value are mmap from the operating system.
        __CFA_DEFAULT_MMAP_START__ = (512 * 1024 + 1),
};

size_t default_mmap_start() __attribute__(( weak )) {
        return __CFA_DEFAULT_MMAP_START__;
} // default_mmap_start

size_t default_heap_expansion() __attribute__(( weak )) {
        return __CFA_DEFAULT_HEAP_EXPANSION__;
} // default_heap_expansion


#ifdef __CFA_DEBUG__
static size_t allocUnfreed;                                                             // running total of allocations minus frees

static void prtUnfreed() {
        if ( allocUnfreed != 0 ) {
                // DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
                char helpText[512];
                int len = snprintf( helpText, sizeof(helpText), "CFA warning (UNIX pid:%ld) : program terminating with %zu(0x%zx) 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
                __cfaabi_bits_write( STDERR_FILENO, helpText, len ); // print debug/nodebug
        } // if
} // prtUnfreed

extern "C" {
        void heapAppStart() {                                                           // called by __cfaabi_appready_startup
                allocUnfreed = 0;
        } // heapAppStart

        void heapAppStop() {                                                            // called by __cfaabi_appready_startdown
                fclose( stdin ); fclose( stdout );
                prtUnfreed();
        } // heapAppStop
} // extern "C"
#endif // __CFA_DEBUG__


// statically allocated variables => zero filled.
static size_t pageSize;                                                                 // architecture pagesize
static size_t heapExpand;                                                               // sbrk advance
static size_t mmapStart;                                                                // cross over point for mmap
static unsigned int maxBucketsUsed;                                             // maximum number of buckets in use


#define SPINLOCK 0
#define LOCKFREE 1
#define BUCKETLOCK SPINLOCK
#if BUCKETLOCK == SPINLOCK
#elif BUCKETLOCK == LOCKFREE
#include <stackLockFree.hfa>
#else
        #error undefined lock type for bucket lock
#endif // LOCKFREE

// Recursive definitions: HeapManager needs size of bucket array and bucket area needs sizeof HeapManager storage.
// Break recusion by hardcoding number of buckets and statically checking number is correct after bucket array defined.
enum { NoBucketSizes = 91 };                                                    // number of buckets sizes

struct HeapManager {
        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 {
                                                                // FreeHeader * home;           // allocated block points back to home locations (must overlay alignment)
                                                                // 2nd low-order bit => zero filled
                                                                void * home;                    // allocated block points back to home locations (must overlay alignment)
                                                                size_t blockSize;               // size for munmap (must overlay alignment)
                                                                #if BUCKETLOCK == SPINLOCK
                                                                Storage * next;                 // freed block points next freed block of same size
                                                                #endif // SPINLOCK
                                                        };
                                                        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
                                                };
                                                #if BUCKETLOCK == LOCKFREE
                                                Link(Storage) next;                             // freed block points next freed block of same size (double-wide)
                                                #endif // LOCKFREE
                                        };
                                } 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;                                     // low-order bits of home/blockSize used for tricks
                                        #endif // __ORDER_BIG_ENDIAN__
                                } fake; // FakeHeader
                        } kind; // Kind
                } header; // Header
                char pad[libAlign() - sizeof( Header )];
                char data[0];                                                                   // storage
        }; // Storage

        static_assert( libAlign() >= sizeof( Storage ), "libAlign() < sizeof( Storage )" );

        struct FreeHeader {
                #if BUCKETLOCK == SPINLOCK
                __spinlock_t lock;                                                              // must be first field for alignment
                Storage * freeList;
                #else
                StackLF(Storage) freeList;
                #endif // BUCKETLOCK
                size_t blockSize;                                                               // size of allocations on this list
        }; // FreeHeader

        // must be first fields for alignment
        __spinlock_t extlock;                                                           // protects allocation-buffer extension
        FreeHeader freeLists[NoBucketSizes];                            // buckets for different allocation sizes

        void * heapBegin;                                                                       // start of heap
        void * heapEnd;                                                                         // logical end of heap
        size_t heapRemaining;                                                           // amount of storage not allocated in the current chunk
}; // HeapManager

#if BUCKETLOCK == LOCKFREE
static inline {
        Link(HeapManager.Storage) * ?`next( HeapManager.Storage * this ) { return &this->header.kind.real.next; }
        void ?{}( HeapManager.FreeHeader & ) {}
        void ^?{}( HeapManager.FreeHeader & ) {}
} // distribution
#endif // LOCKFREE

static inline size_t getKey( const HeapManager.FreeHeader & freeheader ) { return freeheader.blockSize; }


#define FASTLOOKUP
#define __STATISTICS__

// Size of array must harmonize with NoBucketSizes and individual bucket sizes must be multiple of 16.
// Smaller multiples of 16 and powers of 2 are common allocation sizes, so make them generate the minimum required bucket size.
// malloc(0) returns 0p, so no bucket is necessary for 0 bytes returning an address that can be freed.
static const unsigned int bucketSizes[] @= {                    // different bucket sizes
        16 + sizeof(HeapManager.Storage), 32 + sizeof(HeapManager.Storage), 48 + sizeof(HeapManager.Storage), 64 + sizeof(HeapManager.Storage), // 4
        96 + sizeof(HeapManager.Storage), 112 + sizeof(HeapManager.Storage), 128 + sizeof(HeapManager.Storage), // 3
        160, 192, 224, 256 + sizeof(HeapManager.Storage), // 4
        320, 384, 448, 512 + sizeof(HeapManager.Storage), // 4
        640, 768, 896, 1_024 + sizeof(HeapManager.Storage), // 4
        1_536, 2_048 + sizeof(HeapManager.Storage), // 2
        2_560, 3_072, 3_584, 4_096 + sizeof(HeapManager.Storage), // 4
        6_144, 8_192 + sizeof(HeapManager.Storage), // 2
        9_216, 10_240, 11_264, 12_288, 13_312, 14_336, 15_360, 16_384 + sizeof(HeapManager.Storage), // 8
        18_432, 20_480, 22_528, 24_576, 26_624, 28_672, 30_720, 32_768 + sizeof(HeapManager.Storage), // 8
        36_864, 40_960, 45_056, 49_152, 53_248, 57_344, 61_440, 65_536 + sizeof(HeapManager.Storage), // 8
        73_728, 81_920, 90_112, 98_304, 106_496, 114_688, 122_880, 131_072 + sizeof(HeapManager.Storage), // 8
        147_456, 163_840, 180_224, 196_608, 212_992, 229_376, 245_760, 262_144 + sizeof(HeapManager.Storage), // 8
        294_912, 327_680, 360_448, 393_216, 425_984, 458_752, 491_520, 524_288 + sizeof(HeapManager.Storage), // 8
        655_360, 786_432, 917_504, 1_048_576 + sizeof(HeapManager.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(HeapManager.Storage), // 8
        2_621_440, 3_145_728, 3_670_016, 4_194_304 + sizeof(HeapManager.Storage), // 4
};

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

#ifdef FASTLOOKUP
enum { LookupSizes = 65_536 + sizeof(HeapManager.Storage) }; // number of fast lookup sizes
static unsigned char lookup[LookupSizes];                               // O(1) lookup for small sizes
#endif // FASTLOOKUP

static int mmapFd = -1;                                                                 // fake or actual fd for anonymous file
#ifdef __CFA_DEBUG__
static bool heapBoot = 0;                                                               // detect recursion during boot
#endif // __CFA_DEBUG__

// The constructor for heapManager is called explicitly in memory_startup.
static HeapManager heapManager __attribute__(( aligned (128) )) @= {}; // size of cache line to prevent false sharing


#ifdef __STATISTICS__
// Heap statistics counters.
static unsigned int malloc_calls;
static unsigned long long int malloc_storage;
static unsigned int aalloc_calls;
static unsigned long long int aalloc_storage;
static unsigned int calloc_calls;
static unsigned long long int calloc_storage;
static unsigned int memalign_calls;
static unsigned long long int memalign_storage;
static unsigned int amemalign_calls;
static unsigned long long int amemalign_storage;
static unsigned int cmemalign_calls;
static unsigned long long int cmemalign_storage;
static unsigned int resize_calls;
static unsigned long long int resize_storage;
static unsigned int realloc_calls;
static unsigned long long int realloc_storage;
static unsigned int free_calls;
static unsigned long long int free_storage;
static unsigned int mmap_calls;
static unsigned long long int mmap_storage;
static unsigned int munmap_calls;
static unsigned long long int munmap_storage;
static unsigned int sbrk_calls;
static unsigned long long int sbrk_storage;
// Statistics file descriptor (changed by malloc_stats_fd).
static int stat_fd = STDERR_FILENO;                                             // default stderr

// Use "write" because streams may be shutdown when calls are made.
static void printStats() {
        char helpText[1024];
        __cfaabi_bits_print_buffer( STDERR_FILENO, helpText, sizeof(helpText),
                                                                        "\nHeap statistics:\n"
                                                                        "  malloc: calls %u / storage %llu\n"
                                                                        "  aalloc: calls %u / storage %llu\n"
                                                                        "  calloc: calls %u / storage %llu\n"
                                                                        "  memalign: calls %u / storage %llu\n"
                                                                        "  amemalign: calls %u / storage %llu\n"
                                                                        "  cmemalign: calls %u / storage %llu\n"
                                                                        "  resize: calls %u / storage %llu\n"
                                                                        "  realloc: calls %u / storage %llu\n"
                                                                        "  free: calls %u / storage %llu\n"
                                                                        "  mmap: calls %u / storage %llu\n"
                                                                        "  munmap: calls %u / storage %llu\n"
                                                                        "  sbrk: calls %u / storage %llu\n",
                                                                        malloc_calls, malloc_storage,
                                                                        aalloc_calls, aalloc_storage,
                                                                        calloc_calls, calloc_storage,
                                                                        memalign_calls, memalign_storage,
                                                                        amemalign_calls, amemalign_storage,
                                                                        cmemalign_calls, cmemalign_storage,
                                                                        resize_calls, resize_storage,
                                                                        realloc_calls, realloc_storage,
                                                                        free_calls, free_storage,
                                                                        mmap_calls, mmap_storage,
                                                                        munmap_calls, munmap_storage,
                                                                        sbrk_calls, sbrk_storage
                );
} // printStats

static int printStatsXML( FILE * stream ) {                             // see malloc_info
        char helpText[1024];
        int len = snprintf( helpText, sizeof(helpText),
                                                "<malloc version=\"1\">\n"
                                                "<heap nr=\"0\">\n"
                                                "<sizes>\n"
                                                "</sizes>\n"
                                                "<total type=\"malloc\" count=\"%u\" size=\"%llu\"/>\n"
                                                "<total type=\"aalloc\" count=\"%u\" size=\"%llu\"/>\n"
                                                "<total type=\"calloc\" count=\"%u\" size=\"%llu\"/>\n"
                                                "<total type=\"memalign\" count=\"%u\" size=\"%llu\"/>\n"
                                                "<total type=\"amemalign\" count=\"%u\" size=\"%llu\"/>\n"
                                                "<total type=\"cmemalign\" count=\"%u\" size=\"%llu\"/>\n"
                                                "<total type=\"resize\" count=\"%u\" size=\"%llu\"/>\n"
                                                "<total type=\"realloc\" count=\"%u\" size=\"%llu\"/>\n"
                                                "<total type=\"free\" count=\"%u\" size=\"%llu\"/>\n"
                                                "<total type=\"mmap\" count=\"%u\" size=\"%llu\"/>\n"
                                                "<total type=\"munmap\" count=\"%u\" size=\"%llu\"/>\n"
                                                "<total type=\"sbrk\" count=\"%u\" size=\"%llu\"/>\n"
                                                "</malloc>",
                                                malloc_calls, malloc_storage,
                                                aalloc_calls, aalloc_storage,
                                                calloc_calls, calloc_storage,
                                                memalign_calls, memalign_storage,
                                                amemalign_calls, amemalign_storage,
                                                cmemalign_calls, cmemalign_storage,
                                                resize_calls, resize_storage,
                                                realloc_calls, realloc_storage,
                                                free_calls, free_storage,
                                                mmap_calls, mmap_storage,
                                                munmap_calls, munmap_storage,
                                                sbrk_calls, sbrk_storage
                );
        __cfaabi_bits_write( fileno( stream ), helpText, len ); // ensures all bytes written or exit
        return len;
} // printStatsXML
#endif // __STATISTICS__


// thunk problem
size_t Bsearchl( unsigned int key, const unsigned int * vals, size_t dim ) {
        size_t l = 0, m, h = dim;
        while ( l < h ) {
                m = (l + h) / 2;
                if ( (unsigned int &)(vals[m]) < key ) {                // cast away const
                        l = m + 1;
                } else {
                        h = m;
                } // if
        } // while
        return l;
} // Bsearchl


static inline bool setMmapStart( size_t value ) {               // true => mmapped, false => sbrk
  if ( value < pageSize || bucketSizes[NoBucketSizes - 1] < value ) return false;
        mmapStart = value;                                                                      // set global

        // find the closest bucket size less than or equal to the mmapStart size
        maxBucketsUsed = Bsearchl( (unsigned int)mmapStart, bucketSizes, NoBucketSizes ); // binary search
        assert( maxBucketsUsed < NoBucketSizes );                       // subscript failure ?
        assert( mmapStart <= bucketSizes[maxBucketsUsed] ); // search failure ?
        return true;
} // setMmapStart


// <-------+----------------------------------------------------> bsize (bucket size)
// |header |addr
//==================================================================================
//                   align/offset |
// <-----------------<------------+-----------------------------> bsize (bucket size)
//                   |fake-header | addr
#define headerAddr( addr ) ((HeapManager.Storage.Header *)( (char *)addr - sizeof(HeapManager.Storage) ))
#define realHeader( header ) ((HeapManager.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 ( alignment < libAlign() || ! is_pow2( alignment ) ) {
                abort( "Alignment %zu for memory allocation is less than %d and/or not a power of 2.", alignment, libAlign() );
        } // 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( HeapManager.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 __CFA_DEBUG__
                checkAlign( alignment );                                                // check alignment
                #endif // __CFA_DEBUG__
                header = realHeader( header );                                  // backup from fake to real header
        } else {
                alignment = libAlign();                                                 // => no fake header
        } // if
} // fakeHeader


static inline bool headers( const char name[] __attribute__(( unused )), void * addr, HeapManager.Storage.Header *& header, HeapManager.FreeHeader *& freeElem,
                                                        size_t & size, size_t & alignment ) with( heapManager ) {
        header = headerAddr( addr );

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

        #ifdef __CFA_DEBUG__
        checkHeader( addr < heapBegin, name, addr );            // bad low address ?
        #endif // __CFA_DEBUG__

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

        freeElem = (HeapManager.FreeHeader *)((size_t)header->kind.real.home & -3);
        #ifdef __CFA_DEBUG__
        if ( freeElem < &freeLists[0] || &freeLists[NoBucketSizes] <= freeElem ) {
                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 // __CFA_DEBUG__
        size = freeElem->blockSize;
        return false;
} // headers

#ifdef __CFA_DEBUG__
#if __SIZEOF_POINTER__ == 4
#define MASK 0xdeadbeef
#else
#define MASK 0xdeadbeefdeadbeef
#endif
#define STRIDE size_t

static void * Memset( void * addr, STRIDE size ) {              // debug only
        if ( size % sizeof(STRIDE) != 0 ) abort( "Memset() : internal error, size %zd not multiple of %zd.", size, sizeof(STRIDE) );
        if ( (STRIDE)addr % sizeof(STRIDE) != 0 ) abort( "Memset() : internal error, addr %p not multiple of %zd.", addr, sizeof(STRIDE) );

        STRIDE * end = (STRIDE *)addr + size / sizeof(STRIDE);
        for ( STRIDE * p = (STRIDE *)addr; p < end; p += 1 ) *p = MASK;
        return addr;
} // Memset
#endif // __CFA_DEBUG__

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

static inline void * extend( size_t size ) with( heapManager ) {
        lock( extlock __cfaabi_dbg_ctx2 );
        ptrdiff_t rem = heapRemaining - size;
        if ( rem < 0 ) {
                // If the size requested is bigger than the current remaining storage, increase the size of the heap.

                size_t increase = ceiling2( size > heapExpand ? size : heapExpand, pageSize );
                if ( sbrk( increase ) == (void *)-1 ) {                 // failed, no memory ?
                        unlock( extlock );
                        abort( NO_MEMORY_MSG, size );                           // give up
                } // if
                if ( mprotect( (char *)heapEnd + heapRemaining, increase, PROT_READ | PROT_WRITE | PROT_EXEC ) ) {
                        enum { BufferSize = 128 };
                        char helpText[BufferSize];
                        // Do not call strerror( errno ) as it may call malloc.
                        int len = snprintf( helpText, BufferSize, "internal error, extend(), mprotect failure, heapEnd:%p size:%zd, errno:%d.", heapEnd, increase, errno );
                        __cfaabi_bits_write( STDERR_FILENO, helpText, len );
                } // if
                #ifdef __STATISTICS__
                sbrk_calls += 1;
                sbrk_storage += increase;
                #endif // __STATISTICS__
                #ifdef __CFA_DEBUG__
                // Set new memory to garbage so subsequent uninitialized usages might fail.
                //memset( (char *)heapEnd + heapRemaining, '\377', increase );
                Memset( (char *)heapEnd + heapRemaining, increase );
                #endif // __CFA_DEBUG__
                rem = heapRemaining + increase - size;
        } // if

        HeapManager.Storage * block = (HeapManager.Storage *)heapEnd;
        heapRemaining = rem;
        heapEnd = (char *)heapEnd + size;
        unlock( extlock );
        return block;
} // extend


static inline void * doMalloc( size_t size ) with( heapManager ) {
        HeapManager.Storage * block;                                            // pointer to new block of storage

        // 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.

  if ( unlikely( size > ULONG_MAX - sizeof(HeapManager.Storage) ) ) return 0p;
        size_t tsize = size + sizeof(HeapManager.Storage);
        if ( likely( tsize < mmapStart ) ) {                            // small size => sbrk
                size_t posn;
                #ifdef FASTLOOKUP
                if ( tsize < LookupSizes ) posn = lookup[tsize];
                else
                #endif // FASTLOOKUP
                        posn = Bsearchl( (unsigned int)tsize, bucketSizes, (size_t)maxBucketsUsed );
                HeapManager.FreeHeader * freeElem = &freeLists[posn];
                verify( freeElem <= &freeLists[maxBucketsUsed] ); // subscripting error ?
                verify( tsize <= freeElem->blockSize );                 // search failure ?
                tsize = freeElem->blockSize;                                    // total space needed for request

                // Spin until the lock is acquired for this particular size of block.

                #if BUCKETLOCK == SPINLOCK
                lock( freeElem->lock __cfaabi_dbg_ctx2 );
                block = freeElem->freeList;                                             // remove node from stack
                #else
                block = pop( freeElem->freeList );
                #endif // BUCKETLOCK
                if ( unlikely( block == 0p ) ) {                                // no free block ?
                        #if BUCKETLOCK == SPINLOCK
                        unlock( freeElem->lock );
                        #endif // BUCKETLOCK

                        // Freelist for that size was empty, so carve it out of the heap if there's enough left, or get some more
                        // and then carve it off.

                        block = (HeapManager.Storage *)extend( tsize ); // mutual exclusion on call
                #if BUCKETLOCK == SPINLOCK
                } else {
                        freeElem->freeList = block->header.kind.real.next;
                        unlock( freeElem->lock );
                #endif // BUCKETLOCK
                } // if

                block->header.kind.real.home = freeElem;                // pointer back to free list of apropriate size
        } else {                                                                                        // large size => mmap
  if ( unlikely( size > ULONG_MAX - pageSize ) ) return 0p;
                tsize = ceiling2( tsize, pageSize );                    // must be multiple of page size
                #ifdef __STATISTICS__
                __atomic_add_fetch( &mmap_calls, 1, __ATOMIC_SEQ_CST );
                __atomic_add_fetch( &mmap_storage, tsize, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__

                block = (HeapManager.Storage *)mmap( 0, tsize, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANONYMOUS, mmapFd, 0 );
                if ( block == (HeapManager.Storage *)MAP_FAILED ) { // failed ?
                        if ( errno == ENOMEM ) abort( NO_MEMORY_MSG, tsize ); // no memory
                        // Do not call strerror( errno ) as it may call malloc.
                        abort( "(HeapManager &)0x%p.doMalloc() : internal error, mmap failure, size:%zu errno:%d.", &heapManager, tsize, errno );
                } //if
                #ifdef __CFA_DEBUG__
                // Set new memory to garbage so subsequent uninitialized usages might fail.
                //memset( block, '\377', tsize );
                Memset( block, tsize );
                #endif // __CFA_DEBUG__
                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
        verify( ((uintptr_t)addr & (libAlign() - 1)) == 0 ); // minimum alignment ?

        #ifdef __CFA_DEBUG__
        __atomic_add_fetch( &allocUnfreed, tsize, __ATOMIC_SEQ_CST );
        if ( traceHeap() ) {
                enum { BufferSize = 64 };
                char helpText[BufferSize];
                int len = snprintf( helpText, BufferSize, "%p = Malloc( %zu ) (allocated %zu)\n", addr, size, tsize );
                __cfaabi_bits_write( STDERR_FILENO, helpText, len ); // print debug/nodebug
        } // if
        #endif // __CFA_DEBUG__

        return addr;
} // doMalloc


static inline void doFree( void * addr ) with( heapManager ) {
        #ifdef __CFA_DEBUG__
        if ( unlikely( heapManager.heapBegin == 0p ) ) {
                abort( "doFree( %p ) : internal error, called before heap is initialized.", addr );
        } // if
        #endif // __CFA_DEBUG__

        HeapManager.Storage.Header * header;
        HeapManager.FreeHeader * freeElem;
        size_t size, alignment;                                                         // not used (see realloc)

        if ( headers( "free", addr, header, freeElem, size, alignment ) ) { // mmapped ?
                #ifdef __STATISTICS__
                __atomic_add_fetch( &munmap_calls, 1, __ATOMIC_SEQ_CST );
                __atomic_add_fetch( &munmap_storage, size, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__
                if ( munmap( header, size ) == -1 ) {
                        #ifdef __CFA_DEBUG__
                        abort( "Attempt to deallocate storage %p not allocated or with corrupt header.\n"
                                   "Possible cause is invalid pointer.",
                                   addr );
                        #endif // __CFA_DEBUG__
                } // if
        } else {
                #ifdef __CFA_DEBUG__
                // Set free memory to garbage so subsequent usages might fail.
                //memset( ((HeapManager.Storage *)header)->data, '\377', freeElem->blockSize - sizeof( HeapManager.Storage ) );
                Memset( ((HeapManager.Storage *)header)->data, freeElem->blockSize - sizeof( HeapManager.Storage ) );
                #endif // __CFA_DEBUG__

                #ifdef __STATISTICS__
                free_storage += size;
                #endif // __STATISTICS__
                #if BUCKETLOCK == SPINLOCK
                lock( freeElem->lock __cfaabi_dbg_ctx2 );               // acquire spin lock
                header->kind.real.next = freeElem->freeList;    // push on stack
                freeElem->freeList = (HeapManager.Storage *)header;
                unlock( freeElem->lock );                                               // release spin lock
                #else
                push( freeElem->freeList, *(HeapManager.Storage *)header );
                #endif // BUCKETLOCK
        } // if

        #ifdef __CFA_DEBUG__
        __atomic_add_fetch( &allocUnfreed, -size, __ATOMIC_SEQ_CST );
        if ( traceHeap() ) {
                char helpText[64];
                int len = snprintf( helpText, sizeof(helpText), "Free( %p ) size:%zu\n", addr, size );
                __cfaabi_bits_write( STDERR_FILENO, helpText, len ); // print debug/nodebug
        } // if
        #endif // __CFA_DEBUG__
} // doFree


size_t prtFree( HeapManager & manager ) with( manager ) {
        size_t total = 0;
        #ifdef __STATISTICS__
        __cfaabi_bits_acquire();
        __cfaabi_bits_print_nolock( STDERR_FILENO, "\nBin lists (bin size : free blocks on list)\n" );
        #endif // __STATISTICS__
        for ( unsigned int i = 0; i < maxBucketsUsed; i += 1 ) {
                size_t size = freeLists[i].blockSize;
                #ifdef __STATISTICS__
                unsigned int N = 0;
                #endif // __STATISTICS__

                #if BUCKETLOCK == SPINLOCK
                for ( HeapManager.Storage * p = freeLists[i].freeList; p != 0p; p = p->header.kind.real.next ) {
                #else
                        for(;;) {
//              for ( HeapManager.Storage * p = top( freeLists[i].freeList ); p != 0p; p = (p)`next->top ) {
//              for ( HeapManager.Storage * p = top( freeLists[i].freeList ); p != 0p; /* p = getNext( p )->top */) {
//                      HeapManager.Storage * temp = p->header.kind.real.next.top; // FIX ME: direct assignent fails, initialization works`
//                      typeof(p) temp = (( p )`next)->top;                     // FIX ME: direct assignent fails, initialization works`
//                      p = temp;
                #endif // BUCKETLOCK
                        total += size;
                        #ifdef __STATISTICS__
                        N += 1;
                        #endif // __STATISTICS__
                } // for

                #ifdef __STATISTICS__
                __cfaabi_bits_print_nolock( STDERR_FILENO, "%7zu, %-7u  ", size, N );
                if ( (i + 1) % 8 == 0 ) __cfaabi_bits_print_nolock( STDERR_FILENO, "\n" );
                #endif // __STATISTICS__
        } // for
        #ifdef __STATISTICS__
        __cfaabi_bits_print_nolock( STDERR_FILENO, "\ntotal free blocks:%zu\n", total );
        __cfaabi_bits_release();
        #endif // __STATISTICS__
        return (char *)heapEnd - (char *)heapBegin - total;
} // prtFree


static void ?{}( HeapManager & manager ) with( manager ) {
        pageSize = sysconf( _SC_PAGESIZE );

        for ( unsigned int i = 0; i < NoBucketSizes; i += 1 ) { // initialize the free lists
                freeLists[i].blockSize = bucketSizes[i];
        } // for

        #ifdef FASTLOOKUP
        unsigned int idx = 0;
        for ( unsigned int i = 0; i < LookupSizes; i += 1 ) {
                if ( i > bucketSizes[idx] ) idx += 1;
                lookup[i] = idx;
        } // for
        #endif // FASTLOOKUP

        if ( ! setMmapStart( default_mmap_start() ) ) {
                abort( "HeapManager : internal error, mmap start initialization failure." );
        } // if
        heapExpand = default_heap_expansion();

        char * end = (char *)sbrk( 0 );
        heapBegin = heapEnd = sbrk( (char *)ceiling2( (long unsigned int)end, pageSize ) - end ); // move start of heap to multiple of alignment
} // HeapManager


static void ^?{}( HeapManager & ) {
        #ifdef __STATISTICS__
        if ( traceHeapTerm() ) {
                printStats();
                // prtUnfreed() called in heapAppStop()
        } // if
        #endif // __STATISTICS__
} // ~HeapManager


static void memory_startup( void ) __attribute__(( constructor( STARTUP_PRIORITY_MEMORY ) ));
void memory_startup( void ) {
        #ifdef __CFA_DEBUG__
        if ( heapBoot ) {                                                                       // check for recursion during system boot
                // DO NOT USE STREAMS AS THEY MAY BE UNAVAILABLE AT THIS POINT.
                abort( "boot() : internal error, recursively invoked during system boot." );
        } // if
        heapBoot = true;
        #endif // __CFA_DEBUG__

        //verify( heapManager.heapBegin != 0 );
        //heapManager{};
        if ( heapManager.heapBegin == 0p ) heapManager{};       // sanity check
} // memory_startup

static void memory_shutdown( void ) __attribute__(( destructor( STARTUP_PRIORITY_MEMORY ) ));
void memory_shutdown( void ) {
        ^heapManager{};
} // memory_shutdown


static inline void * mallocNoStats( size_t size ) {             // necessary for malloc statistics
        verify( heapManager.heapBegin != 0p );                          // called before memory_startup ?
  if ( unlikely( size ) == 0 ) return 0p;                               // 0 BYTE ALLOCATION RETURNS NULL POINTER

#if __SIZEOF_POINTER__ == 8
        verify( size < ((typeof(size_t))1 << 48) );
#endif // __SIZEOF_POINTER__ == 8
        return doMalloc( size );
} // mallocNoStats


static inline void * callocNoStats( size_t dim, size_t elemSize ) {
        size_t size = dim * elemSize;
  if ( unlikely( size ) == 0 ) return 0p;                               // 0 BYTE ALLOCATION RETURNS NULL POINTER
        char * addr = (char *)mallocNoStats( size );

        HeapManager.Storage.Header * header;
        HeapManager.FreeHeader * freeElem;
        size_t bsize, alignment;
        #ifndef __CFA_DEBUG__
        bool mapped =
        #endif // __CFA_DEBUG__
                headers( "calloc", addr, header, freeElem, bsize, alignment );
        #ifndef __CFA_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 ( ! mapped )
        #endif // __CFA_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;
} // callocNoStats


static inline void * memalignNoStats( size_t alignment, size_t size ) {
  if ( unlikely( size ) == 0 ) return 0p;                               // 0 BYTE ALLOCATION RETURNS NULL POINTER

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

        // if alignment <= default alignment, do normal malloc as two headers are unnecessary
  if ( unlikely( alignment <= libAlign() ) ) return mallocNoStats( size );

        // 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 libAlign() because it is already the minimum alignment
        // add sizeof(Storage) for fake header
        char * addr = (char *)mallocNoStats( size + alignment - libAlign() + sizeof(HeapManager.Storage) );

        // address in the block of the "next" alignment address
        char * user = (char *)ceiling2( (uintptr_t)(addr + sizeof(HeapManager.Storage)), alignment );

        // address of header from malloc
        HeapManager.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
        HeapManager.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;

        return user;
} // memalignNoStats


static inline void * cmemalignNoStats( size_t alignment, size_t dim, size_t elemSize ) {
        size_t size = dim * elemSize;
  if ( unlikely( size ) == 0 ) return 0p;                               // 0 BYTE ALLOCATION RETURNS NULL POINTER
        char * addr = (char *)memalignNoStats( alignment, size );

        HeapManager.Storage.Header * header;
        HeapManager.FreeHeader * freeElem;
        size_t bsize;
        #ifndef __CFA_DEBUG__
        bool mapped =
        #endif // __CFA_DEBUG__
                headers( "cmemalign", addr, header, freeElem, 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 __CFA_DEBUG__
        if ( ! mapped )
        #endif // __CFA_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;
} // cmemalignNoStats


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 ) {
                #ifdef __STATISTICS__
                __atomic_add_fetch( &malloc_calls, 1, __ATOMIC_SEQ_CST );
                __atomic_add_fetch( &malloc_storage, size, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__

                return mallocNoStats( size );
        } // 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 ) {
                size_t size = dim * elemSize;
                #ifdef __STATISTICS__
                __atomic_add_fetch( &aalloc_calls, 1, __ATOMIC_SEQ_CST );
                __atomic_add_fetch( &aalloc_storage, size, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__

                return mallocNoStats( size );
        } // aalloc


        // Same as aalloc() with memory set to zero.
        void * calloc( size_t dim, size_t elemSize ) {
                #ifdef __STATISTICS__
                __atomic_add_fetch( &calloc_calls, 1, __ATOMIC_SEQ_CST );
                __atomic_add_fetch( &calloc_storage, dim * elemSize, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__

                return callocNoStats( dim, elemSize );
        } // calloc


        // Change the size of the memory block pointed to by oaddr to size bytes. The contents are undefined.  If oaddr is
        // 0p, then the call is equivalent to malloc(size), for all values of size; if size is equal to zero, and oaddr is
        // not 0p, then the call is equivalent to free(oaddr). Unless oaddr is 0p, 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 ) {
                #ifdef __STATISTICS__
                __atomic_add_fetch( &resize_calls, 1, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__

                // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
          if ( unlikely( size == 0 ) ) { free( oaddr ); return 0p; } // special cases
          if ( unlikely( oaddr == 0p ) ) {
                        #ifdef __STATISTICS__
                        __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
                        #endif // __STATISTICS__
                        return mallocNoStats( size );
                } // if

                HeapManager.Storage.Header * header;
                HeapManager.FreeHeader * freeElem;
                size_t bsize, oalign;
                headers( "resize", oaddr, header, freeElem, bsize, oalign );
                size_t odsize = dataStorage( bsize, oaddr, header ); // data storage available in bucket

                // same size, DO NOT preserve STICKY PROPERTIES.
                if ( oalign == libAlign() && size <= odsize && odsize <= size * 2 ) { // allow 50% wasted storage for smaller size
                        header->kind.real.blockSize &= -2;                      // no alignment and turn off 0 fill
                        header->kind.real.size = size;                          // reset allocation size
                        return oaddr;
                } // if

                #ifdef __STATISTICS__
                __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__

                // change size, DO NOT preserve STICKY PROPERTIES.
                free( oaddr );
                return mallocNoStats( size );                                   // create new area
        } // 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 ) {
                #ifdef __STATISTICS__
                __atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__

                // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
          if ( unlikely( size == 0 ) ) { free( oaddr ); return 0p; } // special cases
          if ( unlikely( oaddr == 0p ) ) {
                        #ifdef __STATISTICS__
                        __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
                        #endif // __STATISTICS__
                        return mallocNoStats( size );
                } // if

                HeapManager.Storage.Header * header;
                HeapManager.FreeHeader * freeElem;
                size_t bsize, oalign;
                headers( "realloc", oaddr, header, freeElem, 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
                        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

                #ifdef __STATISTICS__
                __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__

                // change size and copy old content to new storage

                void * naddr;
                if ( likely( oalign == libAlign() ) ) {                 // previous request not aligned ?
                        naddr = mallocNoStats( size );                          // create new area
                } else {
                        naddr = memalignNoStats( oalign, size );        // create new aligned area
                } // if

                headers( "realloc", naddr, header, freeElem, bsize, oalign );
                memcpy( naddr, oaddr, min( osize, size ) );             // copy bytes
                free( oaddr );

                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 ) {
                #ifdef __STATISTICS__
                __atomic_add_fetch( &memalign_calls, 1, __ATOMIC_SEQ_CST );
                __atomic_add_fetch( &memalign_storage, size, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__

                return memalignNoStats( alignment, size );
        } // memalign


        // Same as aalloc() with memory alignment.
        void * amemalign( size_t alignment, size_t dim, size_t elemSize ) {
                size_t size = dim * elemSize;
                #ifdef __STATISTICS__
                __atomic_add_fetch( &cmemalign_calls, 1, __ATOMIC_SEQ_CST );
                __atomic_add_fetch( &cmemalign_storage, size, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__

                return memalignNoStats( alignment, size );
        } // amemalign


        // Same as calloc() with memory alignment.
        void * cmemalign( size_t alignment, size_t dim, size_t elemSize ) {
                #ifdef __STATISTICS__
                __atomic_add_fetch( &cmemalign_calls, 1, __ATOMIC_SEQ_CST );
                __atomic_add_fetch( &cmemalign_storage, dim * elemSize, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__

                return cmemalignNoStats( alignment, dim, elemSize );
        } // 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 0p, 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 ( alignment < libAlign() || ! is_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( pageSize, size );
        } // valloc


        // Same as valloc but rounds size to multiple of page size.
        void * pvalloc( size_t size ) {
                return memalign( pageSize, ceiling2( size, pageSize ) );
        } // 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
        // 0p, no operation is performed.
        void free( void * addr ) {
                #ifdef __STATISTICS__
                __atomic_add_fetch( &free_calls, 1, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__

          if ( unlikely( addr == 0p ) ) {                                       // special case
                        // #ifdef __CFA_DEBUG__
                        // if ( traceHeap() ) {
                        //      #define nullmsg "Free( 0x0 ) size:0\n"
                        //      // Do not debug print free( 0p ), as it can cause recursive entry from sprintf.
                        //      __cfaabi_dbg_write( nullmsg, sizeof(nullmsg) - 1 );
                        // } // if
                        // #endif // __CFA_DEBUG__
                        return;
                } // exit

                doFree( addr );
        } // free


        // Returns the alignment of an allocation.
        size_t malloc_alignment( void * addr ) {
          if ( unlikely( addr == 0p ) ) return libAlign();      // minimum alignment
                HeapManager.Storage.Header * header = headerAddr( addr );
                if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
                        return header->kind.fake.alignment & -2;        // remove flag from value
                } else {
                        return libAlign();                                                      // minimum alignment
                } // if
        } // malloc_alignment


        // Set the alignment for an the allocation and return previous alignment or 0 if no alignment.
        size_t $malloc_alignment_set( void * addr, size_t alignment ) {
          if ( unlikely( addr == 0p ) ) return libAlign();      // minimum alignment
                size_t ret;
                HeapManager.Storage.Header * header = headerAddr( addr );
                if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
                        ret = header->kind.fake.alignment & -2;         // remove flag from old value
                        header->kind.fake.alignment = alignment | 1; // add flag to new value
                } else {
                        ret = 0;                                                                        // => no alignment to change
                } // if
                return ret;
        } // $malloc_alignment_set


        // Returns true if the allocation is zero filled, e.g., allocated by calloc().
        bool malloc_zero_fill( void * addr ) {
          if ( unlikely( addr == 0p ) ) return false;           // null allocation is not zero fill
                HeapManager.Storage.Header * header = headerAddr( addr );
                if ( (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

        // Set allocation is zero filled and return previous zero filled.
        bool $malloc_zero_fill_set( void * addr ) {
          if ( unlikely( addr == 0p ) ) return false;           // null allocation is not zero fill
                HeapManager.Storage.Header * header = headerAddr( addr );
                if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
                        header = realHeader( header );                          // backup from fake to real header
                } // if
                bool ret = (header->kind.real.blockSize & 2) != 0; // zero filled ?
                header->kind.real.blockSize |= 2;                               // mark as zero filled
                return ret;
        } // $malloc_zero_fill_set


        // Returns original total allocation size (not bucket size) => array size is dimension * sizeif(T).
        size_t malloc_size( void * addr ) {
          if ( unlikely( addr == 0p ) ) return 0;                       // null allocation has zero size
                HeapManager.Storage.Header * header = headerAddr( addr );
                if ( (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

        // Set allocation size and return previous size.
        size_t $malloc_size_set( void * addr, size_t size ) {
          if ( unlikely( addr == 0p ) ) return 0;                       // null allocation has 0 size
                HeapManager.Storage.Header * header = headerAddr( addr );
                if ( (header->kind.fake.alignment & 1) == 1 ) { // fake header ?
                        header = realHeader( header );                          // backup from fake to real header
                } // if
                size_t ret = header->kind.real.size;
                header->kind.real.size = size;
                return ret;
        } // $malloc_size_set


        // 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 == 0p ) ) return 0;                       // null allocation has 0 size
                HeapManager.Storage.Header * header;
                HeapManager.FreeHeader * freeElem;
                size_t bsize, alignment;

                headers( "malloc_usable_size", addr, header, freeElem, 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( void ) {
                #ifdef __STATISTICS__
                printStats();
                if ( prtFree() ) prtFree( heapManager );
                #endif // __STATISTICS__
        } // malloc_stats


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


        // 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 ) {
                choose( option ) {
                  case M_TOP_PAD:
                        heapExpand = ceiling2( value, pageSize ); 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


        // Exports 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 ) {
          if ( options != 0 ) { errno = EINVAL; return -1; }
                #ifdef __STATISTICS__
                return printStatsXML( stream );
                #else
                return 0;                                                                               // unsupported
                #endif // __STATISTICS__
        } // malloc_info


        // 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 0p;                                                                              // 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
} // extern "C"


// Must have CFA linkage to overload with C linkage realloc.
void * resize( void * oaddr, size_t nalign, size_t size ) {
        #ifdef __STATISTICS__
        __atomic_add_fetch( &resize_calls, 1, __ATOMIC_SEQ_CST );
        #endif // __STATISTICS__

        if ( unlikely( nalign < libAlign() ) ) nalign = libAlign(); // reset alignment to minimum
        #ifdef __CFA_DEBUG__
        else
                checkAlign( nalign );                                                   // check alignment
        #endif // __CFA_DEBUG__

        // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
  if ( unlikely( size == 0 ) ) { free( oaddr ); return 0p; } // special cases
  if ( unlikely( oaddr == 0p ) ) {
                #ifdef __STATISTICS__
                __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__
                return memalignNoStats( nalign, size );
        } // if

        // Attempt to reuse existing alignment.
        HeapManager.Storage.Header * header = headerAddr( oaddr );
        bool isFakeHeader = header->kind.fake.alignment & 1; // old fake header ?
        size_t oalign;
        if ( isFakeHeader ) {
                oalign = header->kind.fake.alignment & -2;              // old alignment
                if ( (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)
                        HeapManager.FreeHeader * freeElem;
                        size_t bsize, oalign;
                        headers( "resize", oaddr, header, freeElem, 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 == libAlign() ) {                             // new alignment also on libAlign => no fake header needed
                return resize( oaddr, size );                                   // duplicate special case checks
        } // if

        #ifdef __STATISTICS__
        __atomic_add_fetch( &resize_storage, size, __ATOMIC_SEQ_CST );
        #endif // __STATISTICS__

        // change size, DO NOT preserve STICKY PROPERTIES.
        free( oaddr );
        return memalignNoStats( nalign, size );                         // create new aligned area
} // resize


void * realloc( void * oaddr, size_t nalign, size_t size ) {
        if ( unlikely( nalign < libAlign() ) ) nalign = libAlign(); // reset alignment to minimum
        #ifdef __CFA_DEBUG__
        else
                checkAlign( nalign );                                                   // check alignment
        #endif // __CFA_DEBUG__

        // If size is equal to 0, either NULL or a pointer suitable to be passed to free() is returned.
  if ( unlikely( size == 0 ) ) { free( oaddr ); return 0p; } // special cases
  if ( unlikely( oaddr == 0p ) ) {
                #ifdef __STATISTICS__
                __atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
                __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
                #endif // __STATISTICS__
                return memalignNoStats( nalign, size );
        } // if

        // Attempt to reuse existing alignment.
        HeapManager.Storage.Header * header = headerAddr( oaddr );
        bool isFakeHeader = header->kind.fake.alignment & 1; // old fake header ?
        size_t oalign;
        if ( isFakeHeader ) {
                oalign = header->kind.fake.alignment & -2;              // old alignment
                if ( (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 alignment and special case checks
                } // if
        } else if ( ! isFakeHeader                                                      // old real header (aligned on libAlign) ?
                                && nalign == libAlign() )                               // new alignment also on libAlign => no fake header needed
                return realloc( oaddr, size );                                  // duplicate alignment and special case checks

        #ifdef __STATISTICS__
        __atomic_add_fetch( &realloc_calls, 1, __ATOMIC_SEQ_CST );
        __atomic_add_fetch( &realloc_storage, size, __ATOMIC_SEQ_CST );
        #endif // __STATISTICS__

        HeapManager.FreeHeader * freeElem;
        size_t bsize;
        headers( "realloc", oaddr, header, freeElem, 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

        headers( "realloc", naddr, header, freeElem, bsize, oalign );
        memcpy( naddr, oaddr, min( osize, size ) );                     // copy bytes
        free( oaddr );

        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


// Local Variables: //
// tab-width: 4 //
// compile-command: "cfa -nodebug -O2 heap.cfa" //
// End: //