]> git.llucax.com Git - software/dgc/cdgc.git/blobdiff - rt/gc/cdgc/gc.d
Convert methods to free functions
[software/dgc/cdgc.git] / rt / gc / cdgc / gc.d
index 73fd1fa25b569b2d9da7da8690f744f9d8dffcd1..dbcd51808b6b6566ae2f912ee0b609842dc73844 100644 (file)
@@ -44,7 +44,7 @@ version = STACKGROWSDOWN;       // growing the stack means subtracting from the
 
 import rt.gc.cdgc.bits: GCBits;
 import rt.gc.cdgc.stats: GCStats, Stats;
-import rt.gc.cdgc.dynarray: DynArray;
+import dynarray = rt.gc.cdgc.dynarray;
 import alloc = rt.gc.cdgc.alloc;
 import opts = rt.gc.cdgc.opts;
 
@@ -92,6 +92,11 @@ package enum BlkAttr : uint
     ALL_BITS = 0b1111_1111
 }
 
+package bool has_pointermap(uint attrs)
+{
+    return !opts.options.conservative && !(attrs & BlkAttr.NO_SCAN);
+}
+
 private
 {
 
@@ -124,2293 +129,2340 @@ private
 }
 
 
-alias GC gc_t;
+enum
+{
+    PAGESIZE =    4096,
+    POOLSIZE =   (4096*256),
+}
 
 
-/* ============================ GC =============================== */
+enum
+{
+    B_16,
+    B_32,
+    B_64,
+    B_128,
+    B_256,
+    B_512,
+    B_1024,
+    B_2048,
+    B_PAGE,             // start of large alloc
+    B_PAGEPLUS,         // continuation of large alloc
+    B_FREE,             // free page
+    B_MAX
+}
 
 
-class GCLock { }                // just a dummy so we can get a global lock
+alias ubyte Bins;
 
 
-const uint GCVERSION = 1;       // increment every time we change interface
-                                // to GC.
+struct List
+{
+    List *next;
+}
 
-Stats stats;
 
-class GC
+struct Range
 {
-    // For passing to debug code
-    static size_t line;
-    static char*  file;
+    void *pbot;
+    void *ptop;
+    int opCmp(in Range other)
+    {
+        if (pbot < other.pbot)
+            return -1;
+        else
+        return cast(int)(pbot > other.pbot);
+    }
+}
 
-    uint gcversion = GCVERSION;
 
-    Gcx *gcx;                   // implementation
-    static ClassInfo gcLock;    // global lock
+const uint binsize[B_MAX] = [ 16,32,64,128,256,512,1024,2048,4096 ];
+const uint notbinsize[B_MAX] = [ ~(16u-1),~(32u-1),~(64u-1),~(128u-1),~(256u-1),
+                                ~(512u-1),~(1024u-1),~(2048u-1),~(4096u-1) ];
 
 
-    void initialize()
-    {
-        opts.parse(cstdlib.getenv("D_GC_OPTS"));
-        gcLock = GCLock.classinfo;
-        gcx = cast(Gcx*) cstdlib.calloc(1, Gcx.sizeof);
-        if (!gcx)
-            onOutOfMemoryError();
-        gcx.initialize();
-        setStackBottom(rt_stackBottom());
-        stats = Stats(this);
-    }
+/* ============================ GC =============================== */
 
 
-    /**
-     *
-     */
-    void enable()
-    {
-        if (!thread_needLock())
-        {
-            assert(gcx.disabled > 0);
-            gcx.disabled--;
-        }
-        else synchronized (gcLock)
-        {
-            assert(gcx.disabled > 0);
-            gcx.disabled--;
-        }
-    }
+class GCLock {} // just a dummy so we can get a global lock
 
 
-    /**
-     *
-     */
-    void disable()
-    {
-        if (!thread_needLock())
-        {
-            gcx.disabled++;
-        }
-        else synchronized (gcLock)
-        {
-            gcx.disabled++;
-        }
-    }
+struct GC
+{
+    // global lock
+    ClassInfo lock;
 
+    void* p_cache;
+    size_t size_cache;
 
-    /**
-     *
-     */
-    uint getAttr(void* p)
-    {
-        if (!p)
-        {
-            return 0;
-        }
+    // !=0 means don't scan stack
+    uint no_stack;
+    bool any_changes;
+    void* stack_bottom;
+    uint inited;
+    /// Turn off collections if > 0
+    int disabled;
 
-        uint go()
-        {
-            Pool* pool = gcx.findPool(p);
-            uint  oldb = 0;
+    /// min(pool.baseAddr)
+    byte *min_addr;
+    /// max(pool.topAddr)
+    byte *max_addr;
 
-            if (pool)
-            {
-                auto biti = cast(size_t)(p - pool.baseAddr) / 16;
+    /// Free list for each size
+    List*[B_MAX] free_list;
 
-                oldb = gcx.getBits(pool, biti);
-            }
-            return oldb;
-        }
+    dynarray.DynArray!(void*) roots;
+    dynarray.DynArray!(Range) ranges;
+    dynarray.DynArray!(Pool) pools;
 
-        if (!thread_needLock())
-        {
-            return go();
-        }
-        else synchronized (gcLock)
-        {
-            return go();
-        }
+    Stats stats;
+}
+
+private GC* gc;
+
+bool Invariant()
+{
+    assert (gc !is null);
+    if (gc.inited) {
+        for (size_t i = 0; i < gc.pools.length; i++) {
+            Pool* pool = gc.pools[i];
+            pool.Invariant();
+            if (i == 0)
+                assert(gc.min_addr == pool.baseAddr);
+            if (i + 1 < gc.pools.length)
+                assert(*pool < gc.pools[i + 1]);
+            else if (i + 1 == gc.pools.length)
+                assert(gc.max_addr == pool.topAddr);
+        }
+
+        gc.roots.Invariant();
+        gc.ranges.Invariant();
+
+        for (size_t i = 0; i < gc.ranges.length; i++) {
+            assert(gc.ranges[i].pbot);
+            assert(gc.ranges[i].ptop);
+            assert(gc.ranges[i].pbot <= gc.ranges[i].ptop);
+        }
+
+        for (size_t i = 0; i < B_PAGE; i++)
+            for (List *list = gc.free_list[i]; list; list = list.next)
+            {
+            }
     }
+    return true;
+}
 
 
-    /**
-     *
-     */
-    uint setAttr(void* p, uint mask)
+/**
+ * Find Pool that pointer is in.
+ * Return null if not in a Pool.
+ * Assume pools is sorted.
+ */
+Pool *findPool(void *p)
+{
+    if (p >= gc.min_addr && p < gc.max_addr)
     {
-        if (!p)
+        if (gc.pools.length == 1)
         {
-            return 0;
+            return gc.pools[0];
         }
 
-        uint go()
+        for (size_t i = 0; i < gc.pools.length; i++)
         {
-            Pool* pool = gcx.findPool(p);
-            uint  oldb = 0;
-
-            if (pool)
+            Pool* pool = gc.pools[i];
+            if (p < pool.topAddr)
             {
-                auto biti = cast(size_t)(p - pool.baseAddr) / 16;
-
-                oldb = gcx.getBits(pool, biti);
-                gcx.setBits(pool, biti, mask);
+                if (pool.baseAddr <= p)
+                    return pool;
+                break;
             }
-            return oldb;
-        }
-
-        if (!thread_needLock())
-        {
-            return go();
-        }
-        else synchronized (gcLock)
-        {
-            return go();
         }
     }
+    return null;
+}
 
 
-    /**
-     *
-     */
-    uint clrAttr(void* p, uint mask)
+/**
+ * Find base address of block containing pointer p.
+ * Returns null if not a gc'd pointer
+ */
+void* findBase(void *p)
+{
+    Pool *pool;
+
+    pool = findPool(p);
+    if (pool)
     {
-        if (!p)
+        size_t offset = cast(size_t)(p - pool.baseAddr);
+        size_t pn = offset / PAGESIZE;
+        Bins   bin = cast(Bins)pool.pagetable[pn];
+
+        // Adjust bit to be at start of allocated memory block
+        if (bin <= B_PAGE)
         {
-            return 0;
+            return pool.baseAddr + (offset & notbinsize[bin]);
         }
-
-        uint go()
+        else if (bin == B_PAGEPLUS)
         {
-            Pool* pool = gcx.findPool(p);
-            uint  oldb = 0;
-
-            if (pool)
+            do
             {
-                auto biti = cast(size_t)(p - pool.baseAddr) / 16;
-
-                oldb = gcx.getBits(pool, biti);
-                gcx.clrBits(pool, biti, mask);
-            }
-            return oldb;
-        }
+                --pn, offset -= PAGESIZE;
+            } while (cast(Bins)pool.pagetable[pn] == B_PAGEPLUS);
 
-        if (!thread_needLock())
-        {
-            return go();
+            return pool.baseAddr + (offset & (offset.max ^ (PAGESIZE-1)));
         }
-        else synchronized (gcLock)
+        else
         {
-            return go();
+            // we are in a B_FREE page
+            return null;
         }
     }
+    return null;
+}
 
 
-    /**
-     *
-     */
-    void *malloc(size_t size, uint bits = 0)
+/**
+ * Find size of pointer p.
+ * Returns 0 if not a gc'd pointer
+ */
+size_t findSize(void *p)
+{
+    Pool*  pool;
+    size_t size = 0;
+
+    pool = findPool(p);
+    if (pool)
     {
-        if (!size)
-        {
-            return null;
-        }
+        size_t pagenum;
+        Bins   bin;
 
-        if (!thread_needLock())
-        {
-            return mallocNoSync(size, bits);
-        }
-        else synchronized (gcLock)
+        pagenum = cast(size_t)(p - pool.baseAddr) / PAGESIZE;
+        bin = cast(Bins)pool.pagetable[pagenum];
+        size = binsize[bin];
+        if (bin == B_PAGE)
         {
-            return mallocNoSync(size, bits);
+            ubyte* pt;
+            size_t i;
+
+            pt = &pool.pagetable[0];
+            for (i = pagenum + 1; i < pool.npages; i++)
+            {
+                if (pt[i] != B_PAGEPLUS)
+                    break;
+            }
+            size = (i - pagenum) * PAGESIZE;
         }
     }
+    return size;
+}
 
 
-    //
-    //
-    //
-    private void *mallocNoSync(size_t size, uint bits = 0)
-    {
-        assert(size != 0);
-
-        stats.malloc_started(size, bits);
-        scope (exit)
-            stats.malloc_finished();
+/**
+ *
+ */
+BlkInfo getInfo(void* p)
+{
+    Pool*   pool;
+    BlkInfo info;
 
-        void *p = null;
-        Bins bin;
+    pool = findPool(p);
+    if (pool)
+    {
+        size_t offset = cast(size_t)(p - pool.baseAddr);
+        size_t pn = offset / PAGESIZE;
+        Bins   bin = cast(Bins)pool.pagetable[pn];
 
-        assert(gcx);
+        ////////////////////////////////////////////////////////////////////
+        // findAddr
+        ////////////////////////////////////////////////////////////////////
 
-        if (opts.options.sentinel)
-            size += SENTINEL_EXTRA;
-
-        // Compute size bin
-        // Cache previous binsize lookup - Dave Fladebo.
-        static size_t lastsize = -1;
-        static Bins lastbin;
-        if (size == lastsize)
-            bin = lastbin;
-        else
+        if (bin <= B_PAGE)
         {
-            bin = gcx.findBin(size);
-            lastsize = size;
-            lastbin = bin;
+            info.base = pool.baseAddr + (offset & notbinsize[bin]);
         }
-
-        if (bin < B_PAGE)
+        else if (bin == B_PAGEPLUS)
         {
-            p = gcx.bucket[bin];
-            if (p is null)
+            do
             {
-                if (!gcx.allocPage(bin) && !gcx.disabled)   // try to find a new page
-                {
-                    if (!thread_needLock())
-                    {
-                        /* Then we haven't locked it yet. Be sure
-                         * and lock for a collection, since a finalizer
-                         * may start a new thread.
-                         */
-                        synchronized (gcLock)
-                        {
-                            gcx.fullcollectshell();
-                        }
-                    }
-                    else if (!gcx.fullcollectshell())       // collect to find a new page
-                    {
-                        //gcx.newPool(1);
-                    }
-                }
-                if (!gcx.bucket[bin] && !gcx.allocPage(bin))
-                {
-                    gcx.newPool(1);         // allocate new pool to find a new page
-                    int result = gcx.allocPage(bin);
-                    if (!result)
-                        onOutOfMemoryError();
-                }
-                p = gcx.bucket[bin];
+                --pn, offset -= PAGESIZE;
             }
+            while (cast(Bins)pool.pagetable[pn] == B_PAGEPLUS);
 
-            // Return next item from free list
-            gcx.bucket[bin] = (cast(List*)p).next;
-            if( !(bits & BlkAttr.NO_SCAN) )
-                memset(p + size, 0, binsize[bin] - size);
-            if (opts.options.mem_stomp)
-                memset(p, 0xF0, size);
-        }
-        else
-        {
-            p = gcx.bigAlloc(size);
-            if (!p)
-                onOutOfMemoryError();
-        }
-        if (opts.options.sentinel) {
-            size -= SENTINEL_EXTRA;
-            p = sentinel_add(p);
-            sentinel_init(p, size);
+            info.base = pool.baseAddr + (offset & (offset.max ^ (PAGESIZE-1)));
+
+            // fix bin for use by size calc below
+            bin = cast(Bins)pool.pagetable[pn];
         }
 
-        if (bits)
+        ////////////////////////////////////////////////////////////////////
+        // findSize
+        ////////////////////////////////////////////////////////////////////
+
+        info.size = binsize[bin];
+        if (bin == B_PAGE)
         {
-            Pool *pool = gcx.findPool(p);
-            assert(pool);
+            ubyte* pt;
+            size_t i;
 
-            gcx.setBits(pool, cast(size_t)(p - pool.baseAddr) / 16, bits);
+            pt = &pool.pagetable[0];
+            for (i = pn + 1; i < pool.npages; i++)
+            {
+                if (pt[i] != B_PAGEPLUS)
+                    break;
+            }
+            info.size = (i - pn) * PAGESIZE;
         }
-        return p;
+
+        ////////////////////////////////////////////////////////////////////
+        // getAttr
+        ////////////////////////////////////////////////////////////////////
+
+        info.attr = getAttr(pool, cast(size_t)(offset / 16));
+        if (!(info.attr & BlkAttr.NO_SCAN))
+            info.size -= (size_t*).sizeof;  // bitmask
     }
+    return info;
+}
 
 
-    /**
-     *
-     */
-    void *calloc(size_t size, uint bits = 0)
+/**
+ * Compute bin for size.
+ */
+static Bins findBin(size_t size)
+{
+    Bins bin;
+    if (size <= 256)
     {
-        if (!size)
+        if (size <= 64)
         {
-            return null;
+            if (size <= 16)
+                bin = B_16;
+            else if (size <= 32)
+                bin = B_32;
+            else
+                bin = B_64;
         }
-
-        if (!thread_needLock())
+        else
+        {
+            if (size <= 128)
+                bin = B_128;
+            else
+                bin = B_256;
+        }
+    }
+    else
+    {
+        if (size <= 1024)
         {
-            return callocNoSync(size, bits);
+            if (size <= 512)
+                bin = B_512;
+            else
+                bin = B_1024;
         }
-        else synchronized (gcLock)
+        else
         {
-            return callocNoSync(size, bits);
+            if (size <= 2048)
+                bin = B_2048;
+            else
+                bin = B_PAGE;
         }
     }
+    return bin;
+}
 
 
-    //
-    //
-    //
-    private void *callocNoSync(size_t size, uint bits = 0)
-    {
-        assert(size != 0);
+/**
+ * Allocate a new pool of at least size bytes.
+ * Sort it into pools.
+ * Mark all memory in the pool as B_FREE.
+ * Return the actual number of bytes reserved or 0 on error.
+ */
+size_t reserveNoSync(size_t size)
+{
+    assert(size != 0);
+    size_t npages = (size + PAGESIZE - 1) / PAGESIZE;
+    Pool*  pool = newPool(npages);
 
-        void *p = mallocNoSync(size, bits);
-        memset(p, 0, size);
-        return p;
-    }
+    if (!pool)
+        return 0;
+    return pool.npages * PAGESIZE;
+}
 
 
-    /**
-     *
-     */
-    void *realloc(void *p, size_t size, uint bits = 0)
+/**
+ * Minimizes physical memory usage by returning free pools to the OS.
+ */
+void minimizeNoSync()
+{
+    size_t n;
+    size_t pn;
+    Pool*  pool;
+
+    for (n = 0; n < gc.pools.length; n++)
     {
-        if (!thread_needLock())
-        {
-            return reallocNoSync(p, size, bits);
-        }
-        else synchronized (gcLock)
+        pool = gc.pools[n];
+        for (pn = 0; pn < pool.npages; pn++)
         {
-            return reallocNoSync(p, size, bits);
+            if (cast(Bins)pool.pagetable[pn] != B_FREE)
+                break;
         }
+        if (pn < pool.npages)
+            continue;
+        pool.Dtor();
+        gc.pools.remove_at(n);
+        n--;
     }
+    gc.min_addr = gc.pools[0].baseAddr;
+    gc.max_addr = gc.pools[gc.pools.length - 1].topAddr;
+}
+
+
+/**
+ * Allocate a chunk of memory that is larger than a page.
+ * Return null if out of memory.
+ */
+void *bigAlloc(size_t size)
+{
+    Pool*  pool;
+    size_t npages;
+    size_t n;
+    size_t pn;
+    size_t freedpages;
+    void*  p;
+    int    state;
 
+    npages = (size + PAGESIZE - 1) / PAGESIZE;
 
-    //
-    //
-    //
-    private void *reallocNoSync(void *p, size_t size, uint bits = 0)
+    for (state = 0; ; )
     {
-        if (!size)
-        {
-            if (p)
-            {
-                freeNoSync(p);
-                p = null;
-            }
-        }
-        else if (!p)
+        // This code could use some refinement when repeatedly
+        // allocating very large arrays.
+
+        for (n = 0; n < gc.pools.length; n++)
         {
-            p = mallocNoSync(size, bits);
+            pool = gc.pools[n];
+            pn = pool.allocPages(npages);
+            if (pn != OPFAIL)
+                goto L1;
         }
-        else
-        {
-            void *p2;
-            size_t psize;
 
-            if (opts.options.sentinel)
+        // Failed
+        switch (state)
+        {
+        case 0:
+            if (gc.disabled)
             {
-                sentinel_Invariant(p);
-                psize = *sentinel_size(p);
-                if (psize != size)
-                {
-                    if (psize)
-                    {
-                        Pool *pool = gcx.findPool(p);
-
-                        if (pool)
-                        {
-                            auto biti = cast(size_t)(p - pool.baseAddr) / 16;
-
-                            if (bits)
-                            {
-                                gcx.clrBits(pool, biti, BlkAttr.ALL_BITS);
-                                gcx.setBits(pool, biti, bits);
-                            }
-                            else
-                            {
-                                bits = gcx.getBits(pool, biti);
-                            }
-                        }
-                    }
-                    p2 = mallocNoSync(size, bits);
-                    if (psize < size)
-                        size = psize;
-                    cstring.memcpy(p2, p, size);
-                    p = p2;
-                }
+                state = 1;
+                continue;
             }
-            else
+            // Try collecting
+            freedpages = fullcollectshell();
+            if (freedpages >= gc.pools.length * ((POOLSIZE / PAGESIZE) / 4))
             {
-                psize = gcx.findSize(p);        // find allocated size
-                if (psize >= PAGESIZE && size >= PAGESIZE)
-                {
-                    auto psz = psize / PAGESIZE;
-                    auto newsz = (size + PAGESIZE - 1) / PAGESIZE;
-                    if (newsz == psz)
-                        return p;
+                state = 1;
+                continue;
+            }
+            // Release empty pools to prevent bloat
+            minimize();
+            // Allocate new pool
+            pool = newPool(npages);
+            if (!pool)
+            {
+                state = 2;
+                continue;
+            }
+            pn = pool.allocPages(npages);
+            assert(pn != OPFAIL);
+            goto L1;
+        case 1:
+            // Release empty pools to prevent bloat
+            minimize();
+            // Allocate new pool
+            pool = newPool(npages);
+            if (!pool)
+                goto Lnomemory;
+            pn = pool.allocPages(npages);
+            assert(pn != OPFAIL);
+            goto L1;
+        case 2:
+            goto Lnomemory;
+        default:
+            assert(false);
+        }
+    }
+
+  L1:
+    pool.pagetable[pn] = B_PAGE;
+    if (npages > 1)
+        memset(&pool.pagetable[pn + 1], B_PAGEPLUS, npages - 1);
+    p = pool.baseAddr + pn * PAGESIZE;
+    memset(cast(char *)p + size, 0, npages * PAGESIZE - size);
+    if (opts.options.mem_stomp)
+        memset(p, 0xF1, size);
+    return p;
+
+  Lnomemory:
+    return null; // let mallocNoSync handle the error
+}
 
-                    auto pool = gcx.findPool(p);
-                    auto pagenum = (p - pool.baseAddr) / PAGESIZE;
 
-                    if (newsz < psz)
-                    {
-                        // Shrink in place
-                        synchronized (gcLock)
-                        {
-                            if (opts.options.mem_stomp)
-                                memset(p + size, 0xF2, psize - size);
-                            pool.freePages(pagenum + newsz, psz - newsz);
-                        }
-                        return p;
-                    }
-                    else if (pagenum + newsz <= pool.npages)
-                    {
-                        // Attempt to expand in place
-                        synchronized (gcLock)
-                        {
-                            for (size_t i = pagenum + psz; 1;)
-                            {
-                                if (i == pagenum + newsz)
-                                {
-                                    if (opts.options.mem_stomp)
-                                        memset(p + psize, 0xF0, size - psize);
-                                    memset(pool.pagetable + pagenum +
-                                            psz, B_PAGEPLUS, newsz - psz);
-                                    return p;
-                                }
-                                if (i == pool.npages)
-                                {
-                                    break;
-                                }
-                                if (pool.pagetable[i] != B_FREE)
-                                    break;
-                                i++;
-                            }
-                        }
-                    }
-                }
-                if (psize < size ||             // if new size is bigger
-                    psize > size * 2)           // or less than half
-                {
-                    if (psize)
-                    {
-                        Pool *pool = gcx.findPool(p);
-
-                        if (pool)
-                        {
-                            auto biti = cast(size_t)(p - pool.baseAddr) / 16;
-
-                            if (bits)
-                            {
-                                gcx.clrBits(pool, biti, BlkAttr.ALL_BITS);
-                                gcx.setBits(pool, biti, bits);
-                            }
-                            else
-                            {
-                                bits = gcx.getBits(pool, biti);
-                            }
-                        }
-                    }
-                    p2 = mallocNoSync(size, bits);
-                    if (psize < size)
-                        size = psize;
-                    cstring.memcpy(p2, p, size);
-                    p = p2;
-                }
-            }
-        }
-        return p;
+/**
+ * Allocate a new pool with at least npages in it.
+ * Sort it into pools.
+ * Return null if failed.
+ */
+Pool *newPool(size_t npages)
+{
+    // Minimum of POOLSIZE
+    if (npages < POOLSIZE/PAGESIZE)
+        npages = POOLSIZE/PAGESIZE;
+    else if (npages > POOLSIZE/PAGESIZE)
+    {
+        // Give us 150% of requested size, so there's room to extend
+        auto n = npages + (npages >> 1);
+        if (n < size_t.max/PAGESIZE)
+            npages = n;
     }
 
-
-    /**
-     * Attempt to in-place enlarge the memory block pointed to by p by at least
-     * minbytes beyond its current capacity, up to a maximum of maxsize.  This
-     * does not attempt to move the memory block (like realloc() does).
-     *
-     * Returns:
-     *  0 if could not extend p,
-     *  total size of entire memory block if successful.
-     */
-    size_t extend(void* p, size_t minsize, size_t maxsize)
+    // Allocate successively larger pools up to 8 megs
+    if (gc.pools.length)
     {
-        if (!thread_needLock())
-        {
-            return extendNoSync(p, minsize, maxsize);
-        }
-        else synchronized (gcLock)
-        {
-            return extendNoSync(p, minsize, maxsize);
-        }
+        size_t n = gc.pools.length;
+        if (n > 8)
+            n = 8;                  // cap pool size at 8 megs
+        n *= (POOLSIZE / PAGESIZE);
+        if (npages < n)
+            npages = n;
     }
 
-
-    //
-    //
-    //
-    private size_t extendNoSync(void* p, size_t minsize, size_t maxsize)
-    in
+    Pool p;
+    p.initialize(npages);
+    if (!p.baseAddr)
     {
-        assert( minsize <= maxsize );
+        p.Dtor();
+        return null;
     }
-    body
+
+    Pool* pool = gc.pools.insert_sorted(p);
+    if (pool)
     {
-        if (opts.options.sentinel)
-        {
-            return 0;
-        }
-        auto psize = gcx.findSize(p);   // find allocated size
-        if (psize < PAGESIZE)
-            return 0;                   // cannot extend buckets
+        gc.min_addr = gc.pools[0].baseAddr;
+        gc.max_addr = gc.pools[gc.pools.length - 1].topAddr;
+    }
+    return pool;
+}
 
-        auto psz = psize / PAGESIZE;
-        auto minsz = (minsize + PAGESIZE - 1) / PAGESIZE;
-        auto maxsz = (maxsize + PAGESIZE - 1) / PAGESIZE;
 
-        auto pool = gcx.findPool(p);
-        auto pagenum = (p - pool.baseAddr) / PAGESIZE;
+/**
+ * Allocate a page of bin's.
+ * Returns:
+ *  0       failed
+ */
+int allocPage(Bins bin)
+{
+    Pool*  pool;
+    size_t n;
+    size_t pn;
+    byte*  p;
+    byte*  ptop;
 
-        size_t sz;
-        for (sz = 0; sz < maxsz; sz++)
-        {
-            auto i = pagenum + psz + sz;
-            if (i == pool.npages)
-                break;
-            if (pool.pagetable[i] != B_FREE)
-            {
-                if (sz < minsz)
-                    return 0;
-                break;
-            }
-        }
-        if (sz < minsz)
-            return 0;
-        if (opts.options.mem_stomp)
-            memset(p + psize, 0xF0, (psz + sz) * PAGESIZE - psize);
-        memset(pool.pagetable + pagenum + psz, B_PAGEPLUS, sz);
-        gcx.p_cache = null;
-        gcx.size_cache = 0;
-        return (psz + sz) * PAGESIZE;
+    for (n = 0; n < gc.pools.length; n++)
+    {
+        pool = gc.pools[n];
+        pn = pool.allocPages(1);
+        if (pn != OPFAIL)
+            goto L1;
     }
+    return 0;               // failed
 
+  L1:
+    pool.pagetable[pn] = cast(ubyte)bin;
 
-    /**
-     *
-     */
-    size_t reserve(size_t size)
-    {
-        if (!size)
-        {
-            return 0;
-        }
+    // Convert page to free list
+    size_t size = binsize[bin];
+    List **b = &gc.free_list[bin];
 
-        if (!thread_needLock())
-        {
-            return reserveNoSync(size);
-        }
-        else synchronized (gcLock)
-        {
-            return reserveNoSync(size);
-        }
+    p = pool.baseAddr + pn * PAGESIZE;
+    ptop = p + PAGESIZE;
+    for (; p < ptop; p += size)
+    {
+        (cast(List *)p).next = *b;
+        *b = cast(List *)p;
     }
+    return 1;
+}
 
 
-    //
-    //
-    //
-    private size_t reserveNoSync(size_t size)
-    {
-        assert(size != 0);
-        assert(gcx);
-
-        return gcx.reserve(size);
-    }
+/**
+ * Marks a range of memory using the conservative bit mask.  Used for
+ * the stack, for the data segment, and additional memory ranges.
+ */
+void mark_conservative(void* pbot, void* ptop)
+{
+    mark(pbot, ptop, PointerMap.init.bits.ptr);
+}
 
 
-    /**
-     *
-     */
-    void free(void *p)
-    {
-        if (!p)
-        {
-            return;
-        }
+/**
+ * Search a range of memory values and mark any pointers into the GC pool.
+ */
+void mark(void *pbot, void *ptop, size_t* pm_bitmask)
+{
+    // TODO: make our own assert because assert uses the GC
+    assert (pbot <= ptop);
 
-        if (!thread_needLock())
-        {
-            return freeNoSync(p);
-        }
-        else synchronized (gcLock)
-        {
-            return freeNoSync(p);
-        }
-    }
+    const BITS_PER_WORD = size_t.sizeof * 8;
 
+    void **p1 = cast(void **)pbot;
+    void **p2 = cast(void **)ptop;
+    size_t pcache = 0;
+    uint changes = 0;
 
-    //
-    //
-    //
-    private void freeNoSync(void *p)
-    {
-        assert (p);
+    size_t type_size = pm_bitmask[0];
+    size_t* pm_bits = pm_bitmask + 1;
 
-        Pool*  pool;
-        size_t pagenum;
-        Bins   bin;
-        size_t biti;
+    //printf("marking range: %p -> %p\n", pbot, ptop);
+    for (; p1 + type_size <= p2; p1 += type_size) {
+        for (size_t n = 0; n < type_size; n++) {
+            // scan bit set for this word
+            if (!(pm_bits[n / BITS_PER_WORD] & (1 << (n % BITS_PER_WORD))))
+                continue;
 
-        // Find which page it is in
-        pool = gcx.findPool(p);
-        if (!pool)                              // if not one of ours
-            return;                             // ignore
-        if (opts.options.sentinel) {
-            sentinel_Invariant(p);
-            p = sentinel_sub(p);
-        }
-        pagenum = cast(size_t)(p - pool.baseAddr) / PAGESIZE;
-        biti = cast(size_t)(p - pool.baseAddr) / 16;
-        gcx.clrBits(pool, biti, BlkAttr.ALL_BITS);
+            void* p = *(p1 + n);
 
-        bin = cast(Bins)pool.pagetable[pagenum];
-        if (bin == B_PAGE)              // if large alloc
-        {
-            // Free pages
-            size_t npages = 1;
-            size_t n = pagenum;
-            while (++n < pool.npages && pool.pagetable[n] == B_PAGEPLUS)
-                npages++;
-            if (opts.options.mem_stomp)
-                memset(p, 0xF2, npages * PAGESIZE);
-            pool.freePages(pagenum, npages);
-        }
-        else
-        {
-            // Add to free list
-            List *list = cast(List*)p;
+            if (p < gc.min_addr || p >= gc.max_addr)
+                continue;
 
-            if (opts.options.mem_stomp)
-                memset(p, 0xF2, binsize[bin]);
+            if ((cast(size_t)p & ~(PAGESIZE-1)) == pcache)
+                continue;
 
-            list.next = gcx.bucket[bin];
-            gcx.bucket[bin] = list;
-        }
-    }
+            Pool* pool = findPool(p);
+            if (pool)
+            {
+                size_t offset = cast(size_t)(p - pool.baseAddr);
+                size_t bit_i;
+                size_t pn = offset / PAGESIZE;
+                Bins   bin = cast(Bins)pool.pagetable[pn];
 
+                // Adjust bit to be at start of allocated memory block
+                if (bin <= B_PAGE)
+                    bit_i = (offset & notbinsize[bin]) >> 4;
+                else if (bin == B_PAGEPLUS)
+                {
+                    do
+                    {
+                        --pn;
+                    }
+                    while (cast(Bins)pool.pagetable[pn] == B_PAGEPLUS);
+                    bit_i = pn * (PAGESIZE / 16);
+                }
+                else
+                {
+                    // Don't mark bits in B_FREE pages
+                    continue;
+                }
 
-    /**
-     * Determine the base address of the block containing p.  If p is not a gc
-     * allocated pointer, return null.
-     */
-    void* addrOf(void *p)
-    {
-        if (!p)
-        {
-            return null;
-        }
+                if (bin >= B_PAGE) // Cache B_PAGE and B_PAGEPLUS lookups
+                    pcache = cast(size_t)p & ~(PAGESIZE-1);
 
-        if (!thread_needLock())
-        {
-            return addrOfNoSync(p);
-        }
-        else synchronized (gcLock)
-        {
-            return addrOfNoSync(p);
+                if (!pool.mark.test(bit_i))
+                {
+                    pool.mark.set(bit_i);
+                    if (!pool.noscan.test(bit_i))
+                    {
+                        pool.scan.set(bit_i);
+                        changes = 1;
+                    }
+                }
+            }
         }
     }
+    if (changes)
+        gc.any_changes = true;
+}
 
+/**
+ * Return number of full pages free'd.
+ */
+size_t fullcollectshell()
+{
+    gc.stats.collection_started();
+    scope (exit)
+        gc.stats.collection_finished();
 
-    //
-    //
-    //
-    void* addrOfNoSync(void *p)
+    // The purpose of the 'shell' is to ensure all the registers
+    // get put on the stack so they'll be scanned
+    void *sp;
+    size_t result;
+    version (GNU)
     {
-        if (!p)
-        {
-            return null;
-        }
-
-        return gcx.findBase(p);
+        gcc.builtins.__builtin_unwind_init();
+        sp = & sp;
     }
-
-
-    /**
-     * Determine the allocated size of pointer p.  If p is an interior pointer
-     * or not a gc allocated pointer, return 0.
-     */
-    size_t sizeOf(void *p)
+    else version(LDC)
     {
-        if (!p)
+        version(X86)
         {
-            return 0;
+            uint eax,ecx,edx,ebx,ebp,esi,edi;
+            asm
+            {
+                mov eax[EBP], EAX      ;
+                mov ecx[EBP], ECX      ;
+                mov edx[EBP], EDX      ;
+                mov ebx[EBP], EBX      ;
+                mov ebp[EBP], EBP      ;
+                mov esi[EBP], ESI      ;
+                mov edi[EBP], EDI      ;
+                mov  sp[EBP], ESP      ;
+            }
         }
-
-        if (!thread_needLock())
+        else version (X86_64)
         {
-            return sizeOfNoSync(p);
+            ulong rax,rbx,rcx,rdx,rbp,rsi,rdi,r8,r9,r10,r11,r12,r13,r14,r15;
+            asm
+            {
+                movq rax[RBP], RAX      ;
+                movq rbx[RBP], RBX      ;
+                movq rcx[RBP], RCX      ;
+                movq rdx[RBP], RDX      ;
+                movq rbp[RBP], RBP      ;
+                movq rsi[RBP], RSI      ;
+                movq rdi[RBP], RDI      ;
+                movq r8 [RBP], R8       ;
+                movq r9 [RBP], R9       ;
+                movq r10[RBP], R10      ;
+                movq r11[RBP], R11      ;
+                movq r12[RBP], R12      ;
+                movq r13[RBP], R13      ;
+                movq r14[RBP], R14      ;
+                movq r15[RBP], R15      ;
+                movq  sp[RBP], RSP      ;
+            }
         }
-        else synchronized (gcLock)
+        else
         {
-            return sizeOfNoSync(p);
+            static assert( false, "Architecture not supported." );
         }
     }
-
-
-    //
-    //
-    //
-    private size_t sizeOfNoSync(void *p)
+    else
+    {
+    asm
+    {
+        pushad              ;
+        mov sp[EBP],ESP     ;
+    }
+    }
+    result = fullcollect(sp);
+    version (GNU)
+    {
+        // nothing to do
+    }
+    else version(LDC)
+    {
+        // nothing to do
+    }
+    else
+    {
+    asm
     {
-        assert (p);
+        popad               ;
+    }
+    }
+    return result;
+}
 
-        if (opts.options.sentinel)
-        {
-            p = sentinel_sub(p);
-            size_t size = gcx.findSize(p);
-
-            // Check for interior pointer
-            // This depends on:
-            // 1) size is a power of 2 for less than PAGESIZE values
-            // 2) base of memory pool is aligned on PAGESIZE boundary
-            if (cast(size_t)p & (size - 1) & (PAGESIZE - 1))
-                size = 0;
-            return size ? size - SENTINEL_EXTRA : 0;
-        }
-        else
-        {
-            if (p == gcx.p_cache)
-                return gcx.size_cache;
 
-            size_t size = gcx.findSize(p);
+/**
+ *
+ */
+size_t fullcollect(void *stackTop)
+{
+    size_t n;
+    Pool*  pool;
 
-            // Check for interior pointer
-            // This depends on:
-            // 1) size is a power of 2 for less than PAGESIZE values
-            // 2) base of memory pool is aligned on PAGESIZE boundary
-            if (cast(size_t)p & (size - 1) & (PAGESIZE - 1))
-                size = 0;
-            else
-            {
-                gcx.p_cache = p;
-                gcx.size_cache = size;
-            }
+    debug(COLLECT_PRINTF) printf("Gcx.fullcollect()\n");
 
-            return size;
-        }
-    }
+    thread_suspendAll();
+    gc.stats.world_stopped();
 
+    gc.p_cache = null;
+    gc.size_cache = 0;
 
-    /**
-     * Determine the base address of the block containing p.  If p is not a gc
-     * allocated pointer, return null.
-     */
-    BlkInfo query(void *p)
+    gc.any_changes = false;
+    for (n = 0; n < gc.pools.length; n++)
     {
-        if (!p)
-        {
-            BlkInfo i;
-            return  i;
-        }
+        pool = gc.pools[n];
+        pool.mark.zero();
+        pool.scan.zero();
+        pool.freebits.zero();
+    }
 
-        if (!thread_needLock())
-        {
-            return queryNoSync(p);
-        }
-        else synchronized (gcLock)
+    // Mark each free entry, so it doesn't get scanned
+    for (n = 0; n < B_PAGE; n++)
+    {
+        for (List *list = gc.free_list[n]; list; list = list.next)
         {
-            return queryNoSync(p);
+            pool = findPool(list);
+            assert(pool);
+            pool.freebits.set(cast(size_t)(cast(byte*)list - pool.baseAddr) / 16);
         }
     }
 
-
-    //
-    //
-    //
-    BlkInfo queryNoSync(void *p)
+    for (n = 0; n < gc.pools.length; n++)
     {
-        assert(p);
+        pool = gc.pools[n];
+        pool.mark.copy(&pool.freebits);
+    }
 
-        return gcx.getInfo(p);
+    void mark_conservative_dg(void* pbot, void* ptop)
+    {
+        mark_conservative(pbot, ptop);
     }
 
+    rt_scanStaticData(&mark_conservative_dg);
 
-    /**
-     * Verify that pointer p:
-     *  1) belongs to this memory pool
-     *  2) points to the start of an allocated piece of memory
-     *  3) is not on a free list
-     */
-    void check(void *p)
+    if (!gc.no_stack)
     {
-        if (!p)
-        {
-            return;
-        }
-
-        if (!thread_needLock())
-        {
-            checkNoSync(p);
-        }
-        else synchronized (gcLock)
-        {
-            checkNoSync(p);
-        }
+        // Scan stacks and registers for each paused thread
+        thread_scanAll(&mark_conservative_dg, stackTop);
     }
 
+    // Scan roots
+    debug(COLLECT_PRINTF) printf("scan roots[]\n");
+    mark_conservative(gc.roots.ptr, gc.roots.ptr + gc.roots.length);
 
-    //
-    //
-    //
-    private void checkNoSync(void *p)
+    // Scan ranges
+    debug(COLLECT_PRINTF) printf("scan ranges[]\n");
+    for (n = 0; n < gc.ranges.length; n++)
     {
-        assert(p);
+        debug(COLLECT_PRINTF) printf("\t%x .. %x\n", gc.ranges[n].pbot, gc.ranges[n].ptop);
+        mark_conservative(gc.ranges[n].pbot, gc.ranges[n].ptop);
+    }
 
-        if (opts.options.sentinel)
-            sentinel_Invariant(p);
-        debug (PTRCHECK)
+    debug(COLLECT_PRINTF) printf("\tscan heap\n");
+    while (gc.any_changes)
+    {
+        gc.any_changes = false;
+        for (n = 0; n < gc.pools.length; n++)
         {
-            Pool*  pool;
-            size_t pagenum;
-            Bins   bin;
-            size_t size;
-
-            if (opts.options.sentinel)
-                p = sentinel_sub(p);
-            pool = gcx.findPool(p);
-            assert(pool);
-            pagenum = cast(size_t)(p - pool.baseAddr) / PAGESIZE;
-            bin = cast(Bins)pool.pagetable[pagenum];
-            assert(bin <= B_PAGE);
-            size = binsize[bin];
-            assert((cast(size_t)p & (size - 1)) == 0);
+            uint *bbase;
+            uint *b;
+            uint *btop;
 
-            debug (PTRCHECK2)
+            pool = gc.pools[n];
+
+            bbase = pool.scan.base();
+            btop = bbase + pool.scan.nwords;
+            for (b = bbase; b < btop;)
             {
-                if (bin < B_PAGE)
+                Bins   bin;
+                size_t pn;
+                size_t u;
+                size_t bitm;
+                byte*  o;
+
+                bitm = *b;
+                if (!bitm)
+                {
+                    b++;
+                    continue;
+                }
+                *b = 0;
+
+                o = pool.baseAddr + (b - bbase) * 32 * 16;
+                if (!(bitm & 0xFFFF))
+                {
+                    bitm >>= 16;
+                    o += 16 * 16;
+                }
+                for (; bitm; o += 16, bitm >>= 1)
                 {
-                    // Check that p is not on a free list
-                    List *list;
+                    if (!(bitm & 1))
+                        continue;
 
-                    for (list = gcx.bucket[bin]; list; list = list.next)
+                    pn = cast(size_t)(o - pool.baseAddr) / PAGESIZE;
+                    bin = cast(Bins)pool.pagetable[pn];
+                    if (bin < B_PAGE) {
+                        if (opts.options.conservative)
+                            mark_conservative(o, o + binsize[bin]);
+                        else {
+                            auto end_of_blk = cast(size_t**)(o +
+                                    binsize[bin] - size_t.sizeof);
+                            size_t* pm_bitmask = *end_of_blk;
+                            mark(o, end_of_blk, pm_bitmask);
+                        }
+                    }
+                    else if (bin == B_PAGE || bin == B_PAGEPLUS)
                     {
-                        assert(cast(void*)list != p);
+                        if (bin == B_PAGEPLUS)
+                        {
+                            while (pool.pagetable[pn - 1] != B_PAGE)
+                                pn--;
+                        }
+                        u = 1;
+                        while (pn + u < pool.npages &&
+                                pool.pagetable[pn + u] == B_PAGEPLUS)
+                            u++;
+
+                        size_t blk_size = u * PAGESIZE;
+                        if (opts.options.conservative)
+                            mark_conservative(o, o + blk_size);
+                        else {
+                            auto end_of_blk = cast(size_t**)(o + blk_size -
+                                    size_t.sizeof);
+                            size_t* pm_bitmask = *end_of_blk;
+                            mark(o, end_of_blk, pm_bitmask);
+                        }
                     }
                 }
             }
         }
     }
 
+    thread_resumeAll();
+    gc.stats.world_started();
 
-    //
-    //
-    //
-    private void setStackBottom(void *p)
+    // Free up everything not marked
+    debug(COLLECT_PRINTF) printf("\tfree'ing\n");
+    size_t freedpages = 0;
+    size_t freed = 0;
+    for (n = 0; n < gc.pools.length; n++)
     {
-        version (STACKGROWSDOWN)
-        {
-            //p = (void *)((uint *)p + 4);
-            if (p > gcx.stackBottom)
-            {
-                gcx.stackBottom = p;
-            }
-        }
-        else
+        pool = gc.pools[n];
+        uint*  bbase = pool.mark.base();
+        size_t pn;
+        for (pn = 0; pn < pool.npages; pn++, bbase += PAGESIZE / (32 * 16))
         {
-            //p = (void *)((uint *)p - 4);
-            if (p < gcx.stackBottom)
-            {
-                gcx.stackBottom = cast(char*)p;
-            }
-        }
-    }
-
+            Bins bin = cast(Bins)pool.pagetable[pn];
 
-    /**
-     * add p to list of roots
-     */
-    void addRoot(void *p)
-    {
-        if (!p)
-        {
-            return;
-        }
+            if (bin < B_PAGE)
+            {
+                auto size = binsize[bin];
+                byte* p = pool.baseAddr + pn * PAGESIZE;
+                byte* ptop = p + PAGESIZE;
+                size_t bit_i = pn * (PAGESIZE/16);
+                size_t bit_stride = size / 16;
 
-        if (!thread_needLock())
-        {
-            if (roots.append(p) is null)
-                onOutOfMemoryError();
-        }
-        else synchronized (gcLock)
-        {
-            if (roots.append(p) is null)
-                onOutOfMemoryError();
-        }
-    }
+version(none) // BUG: doesn't work because freebits() must also be cleared
+{
+                // If free'd entire page
+                if (bbase[0] == 0 && bbase[1] == 0 && bbase[2] == 0 &&
+                        bbase[3] == 0 && bbase[4] == 0 && bbase[5] == 0 &&
+                        bbase[6] == 0 && bbase[7] == 0)
+                {
+                    for (; p < ptop; p += size, bit_i += bit_stride)
+                    {
+                        if (pool.finals.nbits && pool.finals.testClear(bit_i)) {
+                            if (opts.options.sentinel)
+                                rt_finalize(cast(List *)sentinel_add(p), false/*gc.no_stack > 0*/);
+                            else
+                                rt_finalize(cast(List *)p, false/*gc.no_stack > 0*/);
+                        }
+                        clrAttr(pool, bit_i, BlkAttr.ALL_BITS);
 
+                        List *list = cast(List *)p;
 
-    /**
-     * remove p from list of roots
-     */
-    void removeRoot(void *p)
-    {
-        if (!p)
-        {
-            return;
-        }
+                        if (opts.options.mem_stomp)
+                            memset(p, 0xF3, size);
+                    }
+                    pool.pagetable[pn] = B_FREE;
+                    freed += PAGESIZE;
+                    continue;
+                }
+}
+                for (; p < ptop; p += size, bit_i += bit_stride)
+                {
+                    if (!pool.mark.test(bit_i))
+                    {
+                        if (opts.options.sentinel)
+                            sentinel_Invariant(sentinel_add(p));
 
-        bool r;
-        if (!thread_needLock())
-        {
-            r = roots.remove(p);
-        }
-        else synchronized (gcLock)
-        {
-            r = roots.remove(p);
-        }
-        assert (r);
-    }
+                        pool.freebits.set(bit_i);
+                        if (pool.finals.nbits && pool.finals.testClear(bit_i)) {
+                            if (opts.options.sentinel)
+                                rt_finalize(cast(List *)sentinel_add(p), false/*gc.no_stack > 0*/);
+                            else
+                                rt_finalize(cast(List *)p, false/*gc.no_stack > 0*/);
+                        }
+                        clrAttr(pool, bit_i, BlkAttr.ALL_BITS);
 
+                        List *list = cast(List *)p;
 
-    /**
-     * add range to scan for roots
-     */
-    void addRange(void *p, size_t sz)
-    {
-        if (!p || !sz)
-        {
-            return;
-        }
+                        if (opts.options.mem_stomp)
+                            memset(p, 0xF3, size);
 
-        if (!thread_needLock())
-        {
-            if (ranges.append(Range(p, p+sz)) is null)
-                onOutOfMemoryError();
-        }
-        else synchronized (gcLock)
-        {
-            if (ranges.append(Range(p, p+sz)) is null)
-                onOutOfMemoryError();
+                        freed += size;
+                    }
+                }
+            }
+            else if (bin == B_PAGE)
+            {
+                size_t bit_i = pn * (PAGESIZE / 16);
+                if (!pool.mark.test(bit_i))
+                {
+                    byte *p = pool.baseAddr + pn * PAGESIZE;
+                    if (opts.options.sentinel)
+                        sentinel_Invariant(sentinel_add(p));
+                    if (pool.finals.nbits && pool.finals.testClear(bit_i)) {
+                        if (opts.options.sentinel)
+                            rt_finalize(sentinel_add(p), false/*gc.no_stack > 0*/);
+                        else
+                            rt_finalize(p, false/*gc.no_stack > 0*/);
+                    }
+                    clrAttr(pool, bit_i, BlkAttr.ALL_BITS);
+
+                    debug(COLLECT_PRINTF) printf("\tcollecting big %x\n", p);
+                    pool.pagetable[pn] = B_FREE;
+                    freedpages++;
+                    if (opts.options.mem_stomp)
+                        memset(p, 0xF3, PAGESIZE);
+                    while (pn + 1 < pool.npages && pool.pagetable[pn + 1] == B_PAGEPLUS)
+                    {
+                        pn++;
+                        pool.pagetable[pn] = B_FREE;
+                        freedpages++;
+
+                        if (opts.options.mem_stomp)
+                        {
+                            p += PAGESIZE;
+                            memset(p, 0xF3, PAGESIZE);
+                        }
+                    }
+                }
+            }
         }
     }
 
+    // Zero buckets
+    gc.free_list[] = null;
 
-    /**
-     * remove range
-     */
-    void removeRange(void *p)
+    // Free complete pages, rebuild free list
+    debug(COLLECT_PRINTF) printf("\tfree complete pages\n");
+    size_t recoveredpages = 0;
+    for (n = 0; n < gc.pools.length; n++)
     {
-        if (!p)
+        pool = gc.pools[n];
+        for (size_t pn = 0; pn < pool.npages; pn++)
         {
-            return;
-        }
+            Bins   bin = cast(Bins)pool.pagetable[pn];
+            size_t bit_i;
+            size_t u;
 
-        bool r;
-        if (!thread_needLock())
-        {
-            r = ranges.remove(Range(p, null));
-        }
-        else synchronized (gcLock)
-        {
-            r = ranges.remove(Range(p, null));
+            if (bin < B_PAGE)
+            {
+                size_t size = binsize[bin];
+                size_t bit_stride = size / 16;
+                size_t bit_base = pn * (PAGESIZE / 16);
+                size_t bit_top = bit_base + (PAGESIZE / 16);
+                byte*  p;
+
+                bit_i = bit_base;
+                for (; bit_i < bit_top; bit_i += bit_stride)
+                {
+                    if (!pool.freebits.test(bit_i))
+                        goto Lnotfree;
+                }
+                pool.pagetable[pn] = B_FREE;
+                recoveredpages++;
+                continue;
+
+             Lnotfree:
+                p = pool.baseAddr + pn * PAGESIZE;
+                for (u = 0; u < PAGESIZE; u += size)
+                {
+                    bit_i = bit_base + u / 16;
+                    if (pool.freebits.test(bit_i))
+                    {
+                        List *list = cast(List *)(p + u);
+                        // avoid unnecessary writes
+                        if (list.next != gc.free_list[bin])
+                            list.next = gc.free_list[bin];
+                        gc.free_list[bin] = list;
+                    }
+                }
+            }
         }
-        assert (r);
     }
 
+    debug(COLLECT_PRINTF) printf("recovered pages = %d\n", recoveredpages);
+    debug(COLLECT_PRINTF) printf("\tfree'd %u bytes, %u pages from %u pools\n", freed, freedpages, gc.pools.length);
 
-    /**
-     * do full garbage collection
-     */
-    void fullCollect()
-    {
+    return freedpages + recoveredpages;
+}
 
-        if (!thread_needLock())
-        {
-            gcx.fullcollectshell();
-        }
-        else synchronized (gcLock)
-        {
-            gcx.fullcollectshell();
-        }
 
-        version (none)
-        {
-            GCStats stats;
-            getStats(stats);
-        }
+/**
+ *
+ */
+uint getAttr(Pool* pool, size_t bit_i)
+in
+{
+    assert( pool );
+}
+body
+{
+    uint attrs;
 
-    }
+    if (pool.finals.nbits &&
+        pool.finals.test(bit_i))
+        attrs |= BlkAttr.FINALIZE;
+    if (pool.noscan.test(bit_i))
+        attrs |= BlkAttr.NO_SCAN;
+//        if (pool.nomove.nbits &&
+//            pool.nomove.test(bit_i))
+//            attrs |= BlkAttr.NO_MOVE;
+    return attrs;
+}
 
 
-    /**
-     * do full garbage collection ignoring roots
-     */
-    void fullCollectNoStack()
+/**
+ *
+ */
+void setAttr(Pool* pool, size_t bit_i, uint mask)
+in
+{
+    assert( pool );
+}
+body
+{
+    if (mask & BlkAttr.FINALIZE)
     {
-        if (!thread_needLock())
-        {
-            gcx.noStack++;
-            gcx.fullcollectshell();
-            gcx.noStack--;
-        }
-        else synchronized (gcLock)
-        {
-            gcx.noStack++;
-            gcx.fullcollectshell();
-            gcx.noStack--;
-        }
+        if (!pool.finals.nbits)
+            pool.finals.alloc(pool.mark.nbits);
+        pool.finals.set(bit_i);
+    }
+    if (mask & BlkAttr.NO_SCAN)
+    {
+        pool.noscan.set(bit_i);
     }
+//        if (mask & BlkAttr.NO_MOVE)
+//        {
+//            if (!pool.nomove.nbits)
+//                pool.nomove.alloc(pool.mark.nbits);
+//            pool.nomove.set(bit_i);
+//        }
+}
 
 
-    /**
-     * minimize free space usage
-     */
-    void minimize()
+/**
+ *
+ */
+void clrAttr(Pool* pool, size_t bit_i, uint mask)
+in
+{
+    assert( pool );
+}
+body
+{
+    if (mask & BlkAttr.FINALIZE && pool.finals.nbits)
+        pool.finals.clear(bit_i);
+    if (mask & BlkAttr.NO_SCAN)
+        pool.noscan.clear(bit_i);
+//        if (mask & BlkAttr.NO_MOVE && pool.nomove.nbits)
+//            pool.nomove.clear(bit_i);
+}
+
+
+
+void initialize()
+{
+    int dummy;
+    gc.stack_bottom = cast(char*)&dummy;
+    opts.parse(cstdlib.getenv("D_GC_OPTS"));
+    gc.lock = GCLock.classinfo;
+    gc.inited = 1;
+    setStackBottom(rt_stackBottom());
+    gc.stats = Stats(gc);
+}
+
+
+/**
+ *
+ */
+void enable()
+{
+    if (!thread_needLock())
     {
-        if (!thread_needLock())
-        {
-            gcx.minimize();
-        }
-        else synchronized (gcLock)
-        {
-            gcx.minimize();
-        }
+        assert(gc.disabled > 0);
+        gc.disabled--;
+    }
+    else synchronized (gc.lock)
+    {
+        assert(gc.disabled > 0);
+        gc.disabled--;
     }
+}
 
 
-    /**
-     * Retrieve statistics about garbage collection.
-     * Useful for debugging and tuning.
-     */
-    void getStats(out GCStats stats)
+/**
+ *
+ */
+void disable()
+{
+    if (!thread_needLock())
     {
-        if (!thread_needLock())
-        {
-            getStatsNoSync(stats);
-        }
-        else synchronized (gcLock)
-        {
-            getStatsNoSync(stats);
-        }
+        gc.disabled++;
     }
+    else synchronized (gc.lock)
+    {
+        gc.disabled++;
+    }
+}
 
 
-    //
-    //
-    //
-    private void getStatsNoSync(out GCStats stats)
+/**
+ *
+ */
+uint getAttr(void* p)
+{
+    if (!p)
     {
-        size_t psize = 0;
-        size_t usize = 0;
-        size_t flsize = 0;
-
-        size_t n;
-        size_t bsize = 0;
+        return 0;
+    }
 
-        memset(&stats, 0, GCStats.sizeof);
+    uint go()
+    {
+        Pool* pool = findPool(p);
+        uint  old_attrs = 0;
 
-        for (n = 0; n < pools.length; n++)
+        if (pool)
         {
-            Pool* pool = pools[n];
-            psize += pool.npages * PAGESIZE;
-            for (size_t j = 0; j < pool.npages; j++)
-            {
-                Bins bin = cast(Bins)pool.pagetable[j];
-                if (bin == B_FREE)
-                    stats.freeblocks++;
-                else if (bin == B_PAGE)
-                    stats.pageblocks++;
-                else if (bin < B_PAGE)
-                    bsize += PAGESIZE;
-            }
-        }
+            auto bit_i = cast(size_t)(p - pool.baseAddr) / 16;
 
-        for (n = 0; n < B_PAGE; n++)
-        {
-            for (List *list = gcx.bucket[n]; list; list = list.next)
-                flsize += binsize[n];
+            old_attrs = getAttr(pool, bit_i);
         }
+        return old_attrs;
+    }
 
-        usize = bsize - flsize;
-
-        stats.poolsize = psize;
-        stats.usedsize = bsize - flsize;
-        stats.freelistsize = flsize;
+    if (!thread_needLock())
+    {
+        return go();
+    }
+    else synchronized (gc.lock)
+    {
+        return go();
     }
+}
 
-    /******************* weak-reference support *********************/
 
-    // call locked if necessary
-    private T locked(T)(in T delegate() code)
+/**
+ *
+ */
+uint setAttr(void* p, uint mask)
+{
+    if (!p)
     {
-        if (thread_needLock)
-            synchronized(gcLock) return code();
-        else
-           return code();
+        return 0;
     }
 
-    private struct WeakPointer
+    uint go()
     {
-        Object reference;
+        Pool* pool = findPool(p);
+        uint  old_attrs = 0;
 
-        void ondestroy(Object r)
+        if (pool)
         {
-            assert(r is reference);
-            // lock for memory consistency (parallel readers)
-            // also ensures that weakpointerDestroy can be called while another
-            // thread is freeing the reference with "delete"
-            locked!(void)({ reference = null; });
-        }
-    }
+            auto bit_i = cast(size_t)(p - pool.baseAddr) / 16;
 
-    /**
-     * Create a weak pointer to the given object.
-     * Returns a pointer to an opaque struct allocated in C memory.
-     */
-    void* weakpointerCreate( Object r )
-    {
-        if (r)
-       {
-            // must be allocated in C memory
-            // 1. to hide the reference from the GC
-            // 2. the GC doesn't scan delegates added by rt_attachDisposeEvent
-            //    for references
-            auto wp = cast(WeakPointer*)(cstdlib.malloc(WeakPointer.sizeof));
-            if (!wp)
-                onOutOfMemoryError();
-            wp.reference = r;
-            rt_attachDisposeEvent(r, &wp.ondestroy);
-            return wp;
+            old_attrs = getAttr(pool, bit_i);
+            setAttr(pool, bit_i, mask);
         }
-        return null;
+        return old_attrs;
     }
 
-    /**
-     * Destroy a weak pointer returned by weakpointerCreate().
-     * If null is passed, nothing happens.
-     */
-    void weakpointerDestroy( void* p )
+    if (!thread_needLock())
     {
-        if (p)
-       {
-            auto wp = cast(WeakPointer*)p;
-            // must be extra careful about the GC or parallel threads
-            // finalizing the reference at the same time
-            locked!(void)({
-                   if (wp.reference)
-                       rt_detachDisposeEvent(wp.reference, &wp.ondestroy);
-                  });
-            cstdlib.free(wp);
-        }
+        return go();
     }
-
-    /**
-     * Query a weak pointer and return either the object passed to
-     * weakpointerCreate, or null if it was free'd in the meantime.
-     * If null is passed, null is returned.
-     */
-    Object weakpointerGet( void* p )
+    else synchronized (gc.lock)
     {
-        if (p)
-       {
-            // NOTE: could avoid the lock by using Fawzi style GC counters but
-            // that'd require core.sync.Atomic and lots of care about memory
-            // consistency it's an optional optimization see
-            // http://dsource.org/projects/tango/browser/trunk/user/tango/core/Lifetime.d?rev=5100#L158
-            return locked!(Object)({
-                  return (cast(WeakPointer*)p).reference;
-                  });
-            }
+        return go();
     }
 }
 
 
-/* ============================ Gcx =============================== */
-
-enum
-{
-    PAGESIZE =    4096,
-    POOLSIZE =   (4096*256),
-}
-
-
-enum
+/**
+ *
+ */
+uint clrAttr(void* p, uint mask)
 {
-    B_16,
-    B_32,
-    B_64,
-    B_128,
-    B_256,
-    B_512,
-    B_1024,
-    B_2048,
-    B_PAGE,             // start of large alloc
-    B_PAGEPLUS,         // continuation of large alloc
-    B_FREE,             // free page
-    B_MAX
-}
+    if (!p)
+    {
+        return 0;
+    }
 
+    uint go()
+    {
+        Pool* pool = findPool(p);
+        uint  old_attrs = 0;
 
-alias ubyte Bins;
+        if (pool)
+        {
+            auto bit_i = cast(size_t)(p - pool.baseAddr) / 16;
 
+            old_attrs = getAttr(pool, bit_i);
+            clrAttr(pool, bit_i, mask);
+        }
+        return old_attrs;
+    }
 
-struct List
-{
-    List *next;
+    if (!thread_needLock())
+    {
+        return go();
+    }
+    else synchronized (gc.lock)
+    {
+        return go();
+    }
 }
 
 
-struct Range
+/**
+ *
+ */
+void *malloc(size_t size, uint attrs, PointerMap ptrmap)
 {
-    void *pbot;
-    void *ptop;
-    int opCmp(in Range other)
+    if (!size)
     {
-        if (pbot < other.pbot)
-            return -1;
-        else
-        return cast(int)(pbot > other.pbot);
+        return null;
     }
-}
 
+    if (!thread_needLock())
+    {
+        return mallocNoSync(size, attrs, ptrmap.bits.ptr);
+    }
+    else synchronized (gc.lock)
+    {
+        return mallocNoSync(size, attrs, ptrmap.bits.ptr);
+    }
+}
 
-const uint binsize[B_MAX] = [ 16,32,64,128,256,512,1024,2048,4096 ];
-const uint notbinsize[B_MAX] = [ ~(16u-1),~(32u-1),~(64u-1),~(128u-1),~(256u-1),
-                                ~(512u-1),~(1024u-1),~(2048u-1),~(4096u-1) ];
-
-DynArray!(void*) roots;
 
-DynArray!(Range) ranges;
+//
+//
+//
+private void *mallocNoSync(size_t size, uint attrs, size_t* pm_bitmask)
+{
+    assert(size != 0);
 
-DynArray!(Pool) pools;
+    gc.stats.malloc_started(size, attrs, pm_bitmask);
+    scope (exit)
+        gc.stats.malloc_finished(p);
 
+    void *p = null;
+    Bins bin;
 
-/* ============================ Gcx =============================== */
+    if (opts.options.sentinel)
+        size += SENTINEL_EXTRA;
 
+    bool has_pm = has_pointermap(attrs);
+    if (has_pm)
+        size += size_t.sizeof;
 
-struct Gcx
-{
+    // Compute size bin
+    // Cache previous binsize lookup - Dave Fladebo.
+    static size_t lastsize = -1;
+    static Bins lastbin;
+    if (size == lastsize)
+        bin = lastbin;
+    else
+    {
+        bin = findBin(size);
+        lastsize = size;
+        lastbin = bin;
+    }
 
-    void *p_cache;
-    size_t size_cache;
-
-    uint noStack;       // !=0 means don't scan stack
-    uint log;           // turn on logging
-    uint anychanges;
-    void *stackBottom;
-    uint inited;
-    int disabled;       // turn off collections if >0
-
-    byte *minAddr;      // min(baseAddr)
-    byte *maxAddr;      // max(topAddr)
-
-    List *bucket[B_MAX];        // free list for each size
-
-
-    void initialize()
-    {
-        int dummy;
-        (cast(byte*)this)[0 .. Gcx.sizeof] = 0;
-        stackBottom = cast(char*)&dummy;
-        //printf("gcx = %p, self = %x\n", this, self);
-        inited = 1;
-    }
-
-
-    void Invariant() { }
-
-
-    invariant
+    size_t capacity; // to figure out where to store the bitmask
+    if (bin < B_PAGE)
     {
-        if (inited)
+        p = gc.free_list[bin];
+        if (p is null)
         {
-        //printf("Gcx.invariant(): this = %p\n", this);
-            size_t i;
-
-            for (i = 0; i < pools.length; i++)
+            if (!allocPage(bin) && !gc.disabled)   // try to find a new page
             {
-                Pool* pool = pools[i];
-                pool.Invariant();
-                if (i == 0)
-                {
-                    assert(minAddr == pool.baseAddr);
-                }
-                if (i + 1 < pools.length)
+                if (!thread_needLock())
                 {
-                    assert(*pool < pools[i + 1]);
+                    /* Then we haven't locked it yet. Be sure
+                     * and gc.lock for a collection, since a finalizer
+                     * may start a new thread.
+                     */
+                    synchronized (gc.lock)
+                    {
+                        fullcollectshell();
+                    }
                 }
-                else if (i + 1 == pools.length)
+                else if (!fullcollectshell())       // collect to find a new page
                 {
-                    assert(maxAddr == pool.topAddr);
+                    //newPool(1);
                 }
             }
-
-            roots.Invariant();
-            ranges.Invariant();
-
-            for (i = 0; i < ranges.length; i++)
-            {
-                assert(ranges[i].pbot);
-                assert(ranges[i].ptop);
-                assert(ranges[i].pbot <= ranges[i].ptop);
-            }
-
-            for (i = 0; i < B_PAGE; i++)
+            if (!gc.free_list[bin] && !allocPage(bin))
             {
-                for (List *list = bucket[i]; list; list = list.next)
-                {
-                }
+                newPool(1);         // allocate new pool to find a new page
+                int result = allocPage(bin);
+                if (!result)
+                    onOutOfMemoryError();
             }
+            p = gc.free_list[bin];
         }
-    }
-
+        capacity = binsize[bin];
 
-    /**
-     * Find Pool that pointer is in.
-     * Return null if not in a Pool.
-     * Assume pools is sorted.
-     */
-    Pool *findPool(void *p)
+        // Return next item from free list
+        gc.free_list[bin] = (cast(List*)p).next;
+        if (!(attrs & BlkAttr.NO_SCAN))
+            memset(p + size, 0, capacity - size);
+        if (opts.options.mem_stomp)
+            memset(p, 0xF0, size);
+    }
+    else
     {
-        if (p >= minAddr && p < maxAddr)
-        {
-            if (pools.length == 1)
-            {
-                return pools[0];
-            }
+        p = bigAlloc(size);
+        if (!p)
+            onOutOfMemoryError();
+        // Round the size up to the number of pages needed to store it
+        size_t npages = (size + PAGESIZE - 1) / PAGESIZE;
+        capacity = npages * PAGESIZE;
+    }
 
-            for (size_t i = 0; i < pools.length; i++)
-            {
-                Pool* pool = pools[i];
-                if (p < pool.topAddr)
-                {
-                    if (pool.baseAddr <= p)
-                        return pool;
-                    break;
-                }
-            }
-        }
-        return null;
+    // Store the bit mask AFTER SENTINEL_POST
+    // TODO: store it BEFORE, so the bitmask is protected too
+    if (has_pm) {
+        auto end_of_blk = cast(size_t**)(p + capacity - size_t.sizeof);
+        *end_of_blk = pm_bitmask;
+        size -= size_t.sizeof;
     }
 
+    if (opts.options.sentinel) {
+        size -= SENTINEL_EXTRA;
+        p = sentinel_add(p);
+        sentinel_init(p, size);
+    }
 
-    /**
-     * Find base address of block containing pointer p.
-     * Returns null if not a gc'd pointer
-     */
-    void* findBase(void *p)
+    if (attrs)
     {
-        Pool *pool;
+        Pool *pool = findPool(p);
+        assert(pool);
 
-        pool = findPool(p);
-        if (pool)
-        {
-            size_t offset = cast(size_t)(p - pool.baseAddr);
-            size_t pn = offset / PAGESIZE;
-            Bins   bin = cast(Bins)pool.pagetable[pn];
+        setAttr(pool, cast(size_t)(p - pool.baseAddr) / 16, attrs);
+    }
+    return p;
+}
 
-            // Adjust bit to be at start of allocated memory block
-            if (bin <= B_PAGE)
-            {
-                return pool.baseAddr + (offset & notbinsize[bin]);
-            }
-            else if (bin == B_PAGEPLUS)
-            {
-                do
-                {
-                    --pn, offset -= PAGESIZE;
-                } while (cast(Bins)pool.pagetable[pn] == B_PAGEPLUS);
 
-                return pool.baseAddr + (offset & (offset.max ^ (PAGESIZE-1)));
-            }
-            else
-            {
-                // we are in a B_FREE page
-                return null;
-            }
-        }
+/**
+ *
+ */
+void *calloc(size_t size, uint attrs, PointerMap ptrmap)
+{
+    if (!size)
+    {
         return null;
     }
 
-
-    /**
-     * Find size of pointer p.
-     * Returns 0 if not a gc'd pointer
-     */
-    size_t findSize(void *p)
+    if (!thread_needLock())
     {
-        Pool*  pool;
-        size_t size = 0;
+        return callocNoSync(size, attrs, ptrmap.bits.ptr);
+    }
+    else synchronized (gc.lock)
+    {
+        return callocNoSync(size, attrs, ptrmap.bits.ptr);
+    }
+}
 
-        pool = findPool(p);
-        if (pool)
-        {
-            size_t pagenum;
-            Bins   bin;
 
-            pagenum = cast(size_t)(p - pool.baseAddr) / PAGESIZE;
-            bin = cast(Bins)pool.pagetable[pagenum];
-            size = binsize[bin];
-            if (bin == B_PAGE)
-            {
-                ubyte* pt;
-                size_t i;
+//
+//
+//
+private void *callocNoSync(size_t size, uint attrs, size_t* pm_bitmask)
+{
+    assert(size != 0);
 
-                pt = &pool.pagetable[0];
-                for (i = pagenum + 1; i < pool.npages; i++)
-                {
-                    if (pt[i] != B_PAGEPLUS)
-                        break;
-                }
-                size = (i - pagenum) * PAGESIZE;
-            }
-        }
-        return size;
-    }
+    void *p = mallocNoSync(size, attrs, pm_bitmask);
+    memset(p, 0, size);
+    return p;
+}
 
 
-    /**
-     *
-     */
-    BlkInfo getInfo(void* p)
+/**
+ *
+ */
+void *realloc(void *p, size_t size, uint attrs, PointerMap ptrmap)
+{
+    if (!thread_needLock())
     {
-        Pool*   pool;
-        BlkInfo info;
+        return reallocNoSync(p, size, attrs, ptrmap.bits.ptr);
+    }
+    else synchronized (gc.lock)
+    {
+        return reallocNoSync(p, size, attrs, ptrmap.bits.ptr);
+    }
+}
 
-        pool = findPool(p);
-        if (pool)
+
+//
+//
+//
+private void *reallocNoSync(void *p, size_t size, uint attrs,
+        size_t* pm_bitmask)
+{
+    if (!size)
+    {
+        if (p)
         {
-            size_t offset = cast(size_t)(p - pool.baseAddr);
-            size_t pn = offset / PAGESIZE;
-            Bins   bin = cast(Bins)pool.pagetable[pn];
+            freeNoSync(p);
+            p = null;
+        }
+    }
+    else if (!p)
+    {
+        p = mallocNoSync(size, attrs, pm_bitmask);
+    }
+    else
+    {
+        Pool* pool = findPool(p);
+        if (pool is null)
+            return null;
 
-            ////////////////////////////////////////////////////////////////////
-            // findAddr
-            ////////////////////////////////////////////////////////////////////
+        // Set or retrieve attributes as appropriate
+        auto bit_i = cast(size_t)(p - pool.baseAddr) / 16;
+        if (attrs) {
+            clrAttr(pool, bit_i, BlkAttr.ALL_BITS);
+            setAttr(pool, bit_i, attrs);
+        }
+        else
+            attrs = getAttr(pool, bit_i);
+
+        void* blk_base_addr = findBase(p);
+        size_t blk_size = findSize(p);
+        bool has_pm = has_pointermap(attrs);
+        size_t pm_bitmask_size = 0;
+        if (has_pm) {
+            pm_bitmask_size = size_t.sizeof;
+            // Retrieve pointer map bit mask if appropriate
+            if (pm_bitmask is null) {
+                auto end_of_blk = cast(size_t**)(blk_base_addr +
+                        blk_size - size_t.sizeof);
+                pm_bitmask = *end_of_blk;
+            }
+        }
 
-            if (bin <= B_PAGE)
+        if (opts.options.sentinel)
+        {
+            sentinel_Invariant(p);
+            size_t sentinel_stored_size = *sentinel_size(p);
+            if (sentinel_stored_size != size)
             {
-                info.base = pool.baseAddr + (offset & notbinsize[bin]);
+                void* p2 = mallocNoSync(size, attrs, pm_bitmask);
+                if (sentinel_stored_size < size)
+                    size = sentinel_stored_size;
+                cstring.memcpy(p2, p, size);
+                p = p2;
             }
-            else if (bin == B_PAGEPLUS)
+        }
+        else
+        {
+            size += pm_bitmask_size;
+            if (blk_size >= PAGESIZE && size >= PAGESIZE)
             {
-                do
+                auto psz = blk_size / PAGESIZE;
+                auto newsz = (size + PAGESIZE - 1) / PAGESIZE;
+                if (newsz == psz)
+                    return p;
+
+                auto pagenum = (p - pool.baseAddr) / PAGESIZE;
+
+                if (newsz < psz)
+                {
+                    // Shrink in place
+                    synchronized (gc.lock)
+                    {
+                        if (opts.options.mem_stomp)
+                            memset(p + size - pm_bitmask_size, 0xF2,
+                                    blk_size - size - pm_bitmask_size);
+                        pool.freePages(pagenum + newsz, psz - newsz);
+                    }
+                    if (has_pm) {
+                        auto end_of_blk = cast(size_t**)(
+                                blk_base_addr + (PAGESIZE * newsz) -
+                                pm_bitmask_size);
+                        *end_of_blk = pm_bitmask;
+                    }
+                    return p;
+                }
+                else if (pagenum + newsz <= pool.npages)
                 {
-                    --pn, offset -= PAGESIZE;
+                    // Attempt to expand in place
+                    synchronized (gc.lock)
+                    {
+                        for (size_t i = pagenum + psz; 1;)
+                        {
+                            if (i == pagenum + newsz)
+                            {
+                                if (opts.options.mem_stomp)
+                                    memset(p + blk_size - pm_bitmask_size,
+                                            0xF0, size - blk_size
+                                            - pm_bitmask_size);
+                                memset(pool.pagetable + pagenum +
+                                        psz, B_PAGEPLUS, newsz - psz);
+                                if (has_pm) {
+                                    auto end_of_blk = cast(size_t**)(
+                                            blk_base_addr +
+                                            (PAGESIZE * newsz) -
+                                            pm_bitmask_size);
+                                    *end_of_blk = pm_bitmask;
+                                }
+                                return p;
+                            }
+                            if (i == pool.npages)
+                            {
+                                break;
+                            }
+                            if (pool.pagetable[i] != B_FREE)
+                                break;
+                            i++;
+                        }
+                    }
                 }
-                while (cast(Bins)pool.pagetable[pn] == B_PAGEPLUS);
+            }
+            // if new size is bigger or less than half
+            if (blk_size < size || blk_size > size * 2)
+            {
+                size -= pm_bitmask_size;
+                blk_size -= pm_bitmask_size;
+                void* p2 = mallocNoSync(size, attrs, pm_bitmask);
+                if (blk_size < size)
+                    size = blk_size;
+                cstring.memcpy(p2, p, size);
+                p = p2;
+            }
+        }
+    }
+    return p;
+}
 
-                info.base = pool.baseAddr + (offset & (offset.max ^ (PAGESIZE-1)));
 
-                // fix bin for use by size calc below
-                bin = cast(Bins)pool.pagetable[pn];
-            }
+/**
+ * Attempt to in-place enlarge the memory block pointed to by p by at least
+ * minbytes beyond its current capacity, up to a maximum of maxsize.  This
+ * does not attempt to move the memory block (like realloc() does).
+ *
+ * Returns:
+ *  0 if could not extend p,
+ *  total size of entire memory block if successful.
+ */
+size_t extend(void* p, size_t minsize, size_t maxsize)
+{
+    if (!thread_needLock())
+    {
+        return extendNoSync(p, minsize, maxsize);
+    }
+    else synchronized (gc.lock)
+    {
+        return extendNoSync(p, minsize, maxsize);
+    }
+}
 
-            ////////////////////////////////////////////////////////////////////
-            // findSize
-            ////////////////////////////////////////////////////////////////////
 
-            info.size = binsize[bin];
-            if (bin == B_PAGE)
-            {
-                ubyte* pt;
-                size_t i;
+//
+//
+//
+private size_t extendNoSync(void* p, size_t minsize, size_t maxsize)
+in
+{
+    assert( minsize <= maxsize );
+}
+body
+{
+    if (opts.options.sentinel)
+        return 0;
 
-                pt = &pool.pagetable[0];
-                for (i = pn + 1; i < pool.npages; i++)
-                {
-                    if (pt[i] != B_PAGEPLUS)
-                        break;
-                }
-                info.size = (i - pn) * PAGESIZE;
-            }
+    Pool* pool = findPool(p);
+    if (pool is null)
+        return 0;
 
-            ////////////////////////////////////////////////////////////////////
-            // getBits
-            ////////////////////////////////////////////////////////////////////
+    // Retrieve attributes
+    auto bit_i = cast(size_t)(p - pool.baseAddr) / 16;
+    uint attrs = getAttr(pool, bit_i);
 
-            info.attr = getBits(pool, cast(size_t)(offset / 16));
-        }
-        return info;
+    void* blk_base_addr = findBase(p);
+    size_t blk_size = findSize(p);
+    bool has_pm = has_pointermap(attrs);
+    size_t* pm_bitmask = null;
+    size_t pm_bitmask_size = 0;
+    if (has_pm) {
+        pm_bitmask_size = size_t.sizeof;
+        // Retrieve pointer map bit mask
+        auto end_of_blk = cast(size_t**)(blk_base_addr +
+                blk_size - size_t.sizeof);
+        pm_bitmask = *end_of_blk;
+
+        minsize += size_t.sizeof;
+        maxsize += size_t.sizeof;
     }
 
+    if (blk_size < PAGESIZE)
+        return 0; // cannot extend buckets
 
-    /**
-     * Compute bin for size.
-     */
-    static Bins findBin(size_t size)
+    auto psz = blk_size / PAGESIZE;
+    auto minsz = (minsize + PAGESIZE - 1) / PAGESIZE;
+    auto maxsz = (maxsize + PAGESIZE - 1) / PAGESIZE;
+
+    auto pagenum = (p - pool.baseAddr) / PAGESIZE;
+
+    size_t sz;
+    for (sz = 0; sz < maxsz; sz++)
     {
-        Bins bin;
-        if (size <= 256)
+        auto i = pagenum + psz + sz;
+        if (i == pool.npages)
+            break;
+        if (pool.pagetable[i] != B_FREE)
         {
-            if (size <= 64)
-            {
-                if (size <= 16)
-                    bin = B_16;
-                else if (size <= 32)
-                    bin = B_32;
-                else
-                    bin = B_64;
-            }
-            else
-            {
-                if (size <= 128)
-                    bin = B_128;
-                else
-                    bin = B_256;
-            }
-        }
-        else
-        {
-            if (size <= 1024)
-            {
-                if (size <= 512)
-                    bin = B_512;
-                else
-                    bin = B_1024;
-            }
-            else
-            {
-                if (size <= 2048)
-                    bin = B_2048;
-                else
-                    bin = B_PAGE;
-            }
+            if (sz < minsz)
+                return 0;
+            break;
         }
-        return bin;
     }
+    if (sz < minsz)
+        return 0;
 
+    size_t new_size = (psz + sz) * PAGESIZE;
 
-    /**
-     * Allocate a new pool of at least size bytes.
-     * Sort it into pools.
-     * Mark all memory in the pool as B_FREE.
-     * Return the actual number of bytes reserved or 0 on error.
-     */
-    size_t reserve(size_t size)
-    {
-        size_t npages = (size + PAGESIZE - 1) / PAGESIZE;
-        Pool*  pool = newPool(npages);
+    if (opts.options.mem_stomp)
+        memset(p + blk_size - pm_bitmask_size, 0xF0,
+                new_size - blk_size - pm_bitmask_size);
+    memset(pool.pagetable + pagenum + psz, B_PAGEPLUS, sz);
+    gc.p_cache = null;
+    gc.size_cache = 0;
 
-        if (!pool)
-            return 0;
-        return pool.npages * PAGESIZE;
+    if (has_pm) {
+        new_size -= size_t.sizeof;
+        auto end_of_blk = cast(size_t**)(blk_base_addr + new_size);
+        *end_of_blk = pm_bitmask;
     }
+    return new_size;
+}
 
 
-    /**
-     * Minimizes physical memory usage by returning free pools to the OS.
-     */
-    void minimize()
+/**
+ *
+ */
+size_t reserve(size_t size)
+{
+    if (!size)
     {
-        size_t n;
-        size_t pn;
-        Pool*  pool;
+        return 0;
+    }
 
-        for (n = 0; n < pools.length; n++)
-        {
-            pool = pools[n];
-            for (pn = 0; pn < pool.npages; pn++)
-            {
-                if (cast(Bins)pool.pagetable[pn] != B_FREE)
-                    break;
-            }
-            if (pn < pool.npages)
-                continue;
-            pool.Dtor();
-            pools.remove_at(n);
-            n--;
-        }
-        minAddr = pools[0].baseAddr;
-        maxAddr = pools[pools.length - 1].topAddr;
+    if (!thread_needLock())
+    {
+        return reserveNoSync(size);
     }
+    else synchronized (gc.lock)
+    {
+        return reserveNoSync(size);
+    }
+}
 
 
-    /**
-     * Allocate a chunk of memory that is larger than a page.
-     * Return null if out of memory.
-     */
-    void *bigAlloc(size_t size)
+/**
+ *
+ */
+void free(void *p)
+{
+    if (!p)
     {
-        Pool*  pool;
-        size_t npages;
-        size_t n;
-        size_t pn;
-        size_t freedpages;
-        void*  p;
-        int    state;
-
-        npages = (size + PAGESIZE - 1) / PAGESIZE;
+        return;
+    }
 
-        for (state = 0; ; )
-        {
-            // This code could use some refinement when repeatedly
-            // allocating very large arrays.
+    if (!thread_needLock())
+    {
+        return freeNoSync(p);
+    }
+    else synchronized (gc.lock)
+    {
+        return freeNoSync(p);
+    }
+}
 
-            for (n = 0; n < pools.length; n++)
-            {
-                pool = pools[n];
-                pn = pool.allocPages(npages);
-                if (pn != OPFAIL)
-                    goto L1;
-            }
 
-            // Failed
-            switch (state)
-            {
-            case 0:
-                if (disabled)
-                {
-                    state = 1;
-                    continue;
-                }
-                // Try collecting
-                freedpages = fullcollectshell();
-                if (freedpages >= pools.length * ((POOLSIZE / PAGESIZE) / 4))
-                {
-                    state = 1;
-                    continue;
-                }
-                // Release empty pools to prevent bloat
-                minimize();
-                // Allocate new pool
-                pool = newPool(npages);
-                if (!pool)
-                {
-                    state = 2;
-                    continue;
-                }
-                pn = pool.allocPages(npages);
-                assert(pn != OPFAIL);
-                goto L1;
-            case 1:
-                // Release empty pools to prevent bloat
-                minimize();
-                // Allocate new pool
-                pool = newPool(npages);
-                if (!pool)
-                    goto Lnomemory;
-                pn = pool.allocPages(npages);
-                assert(pn != OPFAIL);
-                goto L1;
-            case 2:
-                goto Lnomemory;
-            default:
-                assert(false);
-            }
-        }
+//
+//
+//
+private void freeNoSync(void *p)
+{
+    assert (p);
+
+    Pool*  pool;
+    size_t pagenum;
+    Bins   bin;
+    size_t bit_i;
+
+    // Find which page it is in
+    pool = findPool(p);
+    if (!pool)                              // if not one of ours
+        return;                             // ignore
+    if (opts.options.sentinel) {
+        sentinel_Invariant(p);
+        p = sentinel_sub(p);
+    }
+    pagenum = cast(size_t)(p - pool.baseAddr) / PAGESIZE;
+    bit_i = cast(size_t)(p - pool.baseAddr) / 16;
+    clrAttr(pool, bit_i, BlkAttr.ALL_BITS);
+
+    bin = cast(Bins)pool.pagetable[pagenum];
+    if (bin == B_PAGE)              // if large alloc
+    {
+        // Free pages
+        size_t npages = 1;
+        size_t n = pagenum;
+        while (++n < pool.npages && pool.pagetable[n] == B_PAGEPLUS)
+            npages++;
+        if (opts.options.mem_stomp)
+            memset(p, 0xF2, npages * PAGESIZE);
+        pool.freePages(pagenum, npages);
+    }
+    else
+    {
+        // Add to free list
+        List *list = cast(List*)p;
 
-      L1:
-        pool.pagetable[pn] = B_PAGE;
-        if (npages > 1)
-            memset(&pool.pagetable[pn + 1], B_PAGEPLUS, npages - 1);
-        p = pool.baseAddr + pn * PAGESIZE;
-        memset(cast(char *)p + size, 0, npages * PAGESIZE - size);
         if (opts.options.mem_stomp)
-            memset(p, 0xF1, size);
-        return p;
+            memset(p, 0xF2, binsize[bin]);
 
-      Lnomemory:
-        return null; // let mallocNoSync handle the error
+        list.next = gc.free_list[bin];
+        gc.free_list[bin] = list;
     }
+}
 
 
-    /**
-     * Allocate a new pool with at least npages in it.
-     * Sort it into pools.
-     * Return null if failed.
-     */
-    Pool *newPool(size_t npages)
+/**
+ * Determine the base address of the block containing p.  If p is not a gc
+ * allocated pointer, return null.
+ */
+void* addrOf(void *p)
+{
+    if (!p)
     {
-        // Minimum of POOLSIZE
-        if (npages < POOLSIZE/PAGESIZE)
-            npages = POOLSIZE/PAGESIZE;
-        else if (npages > POOLSIZE/PAGESIZE)
-        {
-            // Give us 150% of requested size, so there's room to extend
-            auto n = npages + (npages >> 1);
-            if (n < size_t.max/PAGESIZE)
-                npages = n;
-        }
-
-        // Allocate successively larger pools up to 8 megs
-        if (pools.length)
-        {
-            size_t n = pools.length;
-            if (n > 8)
-                n = 8;                  // cap pool size at 8 megs
-            n *= (POOLSIZE / PAGESIZE);
-            if (npages < n)
-                npages = n;
-        }
-
-        Pool p;
-        p.initialize(npages);
-        if (!p.baseAddr)
-        {
-            p.Dtor();
-            return null;
-        }
+        return null;
+    }
 
-        Pool* pool = pools.insert_sorted(p);
-        if (pool)
-        {
-            minAddr = pools[0].baseAddr;
-            maxAddr = pools[pools.length - 1].topAddr;
-        }
-        return pool;
+    if (!thread_needLock())
+    {
+        return addrOfNoSync(p);
     }
+    else synchronized (gc.lock)
+    {
+        return addrOfNoSync(p);
+    }
+}
 
 
-    /**
-     * Allocate a page of bin's.
-     * Returns:
-     *  0       failed
-     */
-    int allocPage(Bins bin)
+//
+//
+//
+void* addrOfNoSync(void *p)
+{
+    if (!p)
     {
-        Pool*  pool;
-        size_t n;
-        size_t pn;
-        byte*  p;
-        byte*  ptop;
+        return null;
+    }
 
-        for (n = 0; n < pools.length; n++)
-        {
-            pool = pools[n];
-            pn = pool.allocPages(1);
-            if (pn != OPFAIL)
-                goto L1;
-        }
-        return 0;               // failed
+    return findBase(p);
+}
 
-      L1:
-        pool.pagetable[pn] = cast(ubyte)bin;
 
-        // Convert page to free list
-        size_t size = binsize[bin];
-        List **b = &bucket[bin];
+/**
+ * Determine the allocated size of pointer p.  If p is an interior pointer
+ * or not a gc allocated pointer, return 0.
+ */
+size_t sizeOf(void *p)
+{
+    if (!p)
+    {
+        return 0;
+    }
 
-        p = pool.baseAddr + pn * PAGESIZE;
-        ptop = p + PAGESIZE;
-        for (; p < ptop; p += size)
-        {
-            (cast(List *)p).next = *b;
-            *b = cast(List *)p;
-        }
-        return 1;
+    if (!thread_needLock())
+    {
+        return sizeOfNoSync(p);
+    }
+    else synchronized (gc.lock)
+    {
+        return sizeOfNoSync(p);
     }
+}
 
 
-    /**
-     * Search a range of memory values and mark any pointers into the GC pool.
-     */
-    void mark(void *pbot, void *ptop)
-    {
-        void **p1 = cast(void **)pbot;
-        void **p2 = cast(void **)ptop;
-        size_t pcache = 0;
-        uint changes = 0;
+//
+//
+//
+private size_t sizeOfNoSync(void *p)
+{
+    assert (p);
 
-        //printf("marking range: %p -> %p\n", pbot, ptop);
-        for (; p1 < p2; p1++)
-        {
-            Pool *pool;
-            byte *p = cast(byte *)(*p1);
+    if (opts.options.sentinel)
+        p = sentinel_sub(p);
 
-            if (p >= minAddr && p < maxAddr)
-            {
-                if ((cast(size_t)p & ~(PAGESIZE-1)) == pcache)
-                    continue;
+    Pool* pool = findPool(p);
+    if (pool is null)
+        return 0;
 
-                pool = findPool(p);
-                if (pool)
-                {
-                    size_t offset = cast(size_t)(p - pool.baseAddr);
-                    size_t biti;
-                    size_t pn = offset / PAGESIZE;
-                    Bins   bin = cast(Bins)pool.pagetable[pn];
-
-                    // Adjust bit to be at start of allocated memory block
-                    if (bin <= B_PAGE)
-                        biti = (offset & notbinsize[bin]) >> 4;
-                    else if (bin == B_PAGEPLUS)
-                    {
-                        do
-                        {
-                            --pn;
-                        }
-                        while (cast(Bins)pool.pagetable[pn] == B_PAGEPLUS);
-                        biti = pn * (PAGESIZE / 16);
-                    }
-                    else
-                    {
-                        // Don't mark bits in B_FREE pages
-                        continue;
-                    }
+    auto biti = cast(size_t)(p - pool.baseAddr) / 16;
+    uint attrs = getAttr(pool, biti);
 
-                    if (bin >= B_PAGE) // Cache B_PAGE and B_PAGEPLUS lookups
-                        pcache = cast(size_t)p & ~(PAGESIZE-1);
+    size_t size = findSize(p);
+    size_t pm_bitmask_size = 0;
+    if (has_pointermap(attrs))
+        pm_bitmask_size = size_t.sizeof;
 
-                    if (!pool.mark.test(biti))
-                    {
-                        pool.mark.set(biti);
-                        if (!pool.noscan.test(biti))
-                        {
-                            pool.scan.set(biti);
-                            changes = 1;
-                        }
-                    }
-                }
-            }
-        }
-        anychanges |= changes;
+    if (opts.options.sentinel) {
+        // Check for interior pointer
+        // This depends on:
+        // 1) size is a power of 2 for less than PAGESIZE values
+        // 2) base of memory pool is aligned on PAGESIZE boundary
+        if (cast(size_t)p & (size - 1) & (PAGESIZE - 1))
+            return 0;
+        return size - SENTINEL_EXTRA - pm_bitmask_size;
     }
+    else {
+        if (p == gc.p_cache)
+            return gc.size_cache;
 
-    /**
-     * Return number of full pages free'd.
-     */
-    size_t fullcollectshell()
-    {
-        stats.collection_started();
-        scope (exit)
-            stats.collection_finished();
+        // Check for interior pointer
+        // This depends on:
+        // 1) size is a power of 2 for less than PAGESIZE values
+        // 2) base of memory pool is aligned on PAGESIZE boundary
+        if (cast(size_t)p & (size - 1) & (PAGESIZE - 1))
+            return 0;
 
-        // The purpose of the 'shell' is to ensure all the registers
-        // get put on the stack so they'll be scanned
-        void *sp;
-        size_t result;
-        version (GNU)
-        {
-            gcc.builtins.__builtin_unwind_init();
-            sp = & sp;
-        }
-        else version(LDC)
-        {
-            version(X86)
-            {
-                uint eax,ecx,edx,ebx,ebp,esi,edi;
-                asm
-                {
-                    mov eax[EBP], EAX      ;
-                    mov ecx[EBP], ECX      ;
-                    mov edx[EBP], EDX      ;
-                    mov ebx[EBP], EBX      ;
-                    mov ebp[EBP], EBP      ;
-                    mov esi[EBP], ESI      ;
-                    mov edi[EBP], EDI      ;
-                    mov  sp[EBP], ESP      ;
-                }
-            }
-            else version (X86_64)
-            {
-                ulong rax,rbx,rcx,rdx,rbp,rsi,rdi,r8,r9,r10,r11,r12,r13,r14,r15;
-                asm
-                {
-                    movq rax[RBP], RAX      ;
-                    movq rbx[RBP], RBX      ;
-                    movq rcx[RBP], RCX      ;
-                    movq rdx[RBP], RDX      ;
-                    movq rbp[RBP], RBP      ;
-                    movq rsi[RBP], RSI      ;
-                    movq rdi[RBP], RDI      ;
-                    movq r8 [RBP], R8       ;
-                    movq r9 [RBP], R9       ;
-                    movq r10[RBP], R10      ;
-                    movq r11[RBP], R11      ;
-                    movq r12[RBP], R12      ;
-                    movq r13[RBP], R13      ;
-                    movq r14[RBP], R14      ;
-                    movq r15[RBP], R15      ;
-                    movq  sp[RBP], RSP      ;
-                }
-            }
-            else
-            {
-                static assert( false, "Architecture not supported." );
-            }
-        }
-        else
-        {
-        asm
-        {
-            pushad              ;
-            mov sp[EBP],ESP     ;
-        }
-        }
-        result = fullcollect(sp);
-        version (GNU)
-        {
-            // nothing to do
-        }
-        else version(LDC)
-        {
-            // nothing to do
-        }
-        else
-        {
-        asm
-        {
-            popad               ;
-        }
-        }
-        return result;
+        gc.p_cache = p;
+        gc.size_cache = size - pm_bitmask_size;
+
+        return gc.size_cache;
     }
+}
 
 
-    /**
-     *
-     */
-    size_t fullcollect(void *stackTop)
+/**
+ * Determine the base address of the block containing p.  If p is not a gc
+ * allocated pointer, return null.
+ */
+BlkInfo query(void *p)
+{
+    if (!p)
     {
-        size_t n;
-        Pool*  pool;
+        BlkInfo i;
+        return  i;
+    }
 
-        debug(COLLECT_PRINTF) printf("Gcx.fullcollect()\n");
+    if (!thread_needLock())
+    {
+        return queryNoSync(p);
+    }
+    else synchronized (gc.lock)
+    {
+        return queryNoSync(p);
+    }
+}
 
-        thread_suspendAll();
-        stats.world_stopped();
 
-        p_cache = null;
-        size_cache = 0;
+//
+//
+//
+BlkInfo queryNoSync(void *p)
+{
+    assert(p);
 
-        anychanges = 0;
-        for (n = 0; n < pools.length; n++)
-        {
-            pool = pools[n];
-            pool.mark.zero();
-            pool.scan.zero();
-            pool.freebits.zero();
-        }
+    return getInfo(p);
+}
 
-        // Mark each free entry, so it doesn't get scanned
-        for (n = 0; n < B_PAGE; n++)
-        {
-            for (List *list = bucket[n]; list; list = list.next)
-            {
-                pool = findPool(list);
-                assert(pool);
-                pool.freebits.set(cast(size_t)(cast(byte*)list - pool.baseAddr) / 16);
-            }
-        }
 
-        for (n = 0; n < pools.length; n++)
-        {
-            pool = pools[n];
-            pool.mark.copy(&pool.freebits);
-        }
+/**
+ * Verify that pointer p:
+ *  1) belongs to this memory pool
+ *  2) points to the start of an allocated piece of memory
+ *  3) is not on a free list
+ */
+void check(void *p)
+{
+    if (!p)
+    {
+        return;
+    }
 
-        rt_scanStaticData( &mark );
+    if (!thread_needLock())
+    {
+        checkNoSync(p);
+    }
+    else synchronized (gc.lock)
+    {
+        checkNoSync(p);
+    }
+}
 
-        if (!noStack)
-        {
-            // Scan stacks and registers for each paused thread
-            thread_scanAll( &mark, stackTop );
-        }
 
-        // Scan roots
-        debug(COLLECT_PRINTF) printf("scan roots[]\n");
-        mark(roots.ptr, roots.ptr + roots.length);
+//
+//
+//
+private void checkNoSync(void *p)
+{
+    assert(p);
 
-        // Scan ranges
-        debug(COLLECT_PRINTF) printf("scan ranges[]\n");
-        //log++;
-        for (n = 0; n < ranges.length; n++)
-        {
-            debug(COLLECT_PRINTF) printf("\t%x .. %x\n", ranges[n].pbot, ranges[n].ptop);
-            mark(ranges[n].pbot, ranges[n].ptop);
-        }
-        //log--;
+    if (opts.options.sentinel)
+        sentinel_Invariant(p);
+    debug (PTRCHECK)
+    {
+        Pool*  pool;
+        size_t pagenum;
+        Bins   bin;
+        size_t size;
+
+        if (opts.options.sentinel)
+            p = sentinel_sub(p);
+        pool = findPool(p);
+        assert(pool);
+        pagenum = cast(size_t)(p - pool.baseAddr) / PAGESIZE;
+        bin = cast(Bins)pool.pagetable[pagenum];
+        assert(bin <= B_PAGE);
+        size = binsize[bin];
+        assert((cast(size_t)p & (size - 1)) == 0);
 
-        debug(COLLECT_PRINTF) printf("\tscan heap\n");
-        while (anychanges)
+        debug (PTRCHECK2)
         {
-            anychanges = 0;
-            for (n = 0; n < pools.length; n++)
+            if (bin < B_PAGE)
             {
-                uint *bbase;
-                uint *b;
-                uint *btop;
+                // Check that p is not on a free list
+                List *list;
 
-                pool = pools[n];
-
-                bbase = pool.scan.base();
-                btop = bbase + pool.scan.nwords;
-                for (b = bbase; b < btop;)
+                for (list = gc.free_list[bin]; list; list = list.next)
                 {
-                    Bins   bin;
-                    size_t pn;
-                    size_t u;
-                    size_t bitm;
-                    byte*  o;
-
-                    bitm = *b;
-                    if (!bitm)
-                    {
-                        b++;
-                        continue;
-                    }
-                    *b = 0;
-
-                    o = pool.baseAddr + (b - bbase) * 32 * 16;
-                    if (!(bitm & 0xFFFF))
-                    {
-                        bitm >>= 16;
-                        o += 16 * 16;
-                    }
-                    for (; bitm; o += 16, bitm >>= 1)
-                    {
-                        if (!(bitm & 1))
-                            continue;
-
-                        pn = cast(size_t)(o - pool.baseAddr) / PAGESIZE;
-                        bin = cast(Bins)pool.pagetable[pn];
-                        if (bin < B_PAGE)
-                        {
-                            mark(o, o + binsize[bin]);
-                        }
-                        else if (bin == B_PAGE || bin == B_PAGEPLUS)
-                        {
-                            if (bin == B_PAGEPLUS)
-                            {
-                                while (pool.pagetable[pn - 1] != B_PAGE)
-                                    pn--;
-                            }
-                            u = 1;
-                            while (pn + u < pool.npages && pool.pagetable[pn + u] == B_PAGEPLUS)
-                                u++;
-                            mark(o, o + u * PAGESIZE);
-                        }
-                    }
+                    assert(cast(void*)list != p);
                 }
             }
         }
+    }
+}
 
-        thread_resumeAll();
-        stats.world_started();
 
-        // Free up everything not marked
-        debug(COLLECT_PRINTF) printf("\tfree'ing\n");
-        size_t freedpages = 0;
-        size_t freed = 0;
-        for (n = 0; n < pools.length; n++)
+//
+//
+//
+private void setStackBottom(void *p)
+{
+    version (STACKGROWSDOWN)
+    {
+        //p = (void *)((uint *)p + 4);
+        if (p > gc.stack_bottom)
         {
-            pool = pools[n];
-            uint*  bbase = pool.mark.base();
-            size_t pn;
-            for (pn = 0; pn < pool.npages; pn++, bbase += PAGESIZE / (32 * 16))
-            {
-                Bins bin = cast(Bins)pool.pagetable[pn];
+            gc.stack_bottom = p;
+        }
+    }
+    else
+    {
+        //p = (void *)((uint *)p - 4);
+        if (p < gc.stack_bottom)
+        {
+            gc.stack_bottom = cast(char*)p;
+        }
+    }
+}
+
+
+/**
+ * add p to list of roots
+ */
+void addRoot(void *p)
+{
+    if (!p)
+    {
+        return;
+    }
+
+    if (!thread_needLock())
+    {
+        if (gc.roots.append(p) is null)
+            onOutOfMemoryError();
+    }
+    else synchronized (gc.lock)
+    {
+        if (gc.roots.append(p) is null)
+            onOutOfMemoryError();
+    }
+}
 
-                if (bin < B_PAGE)
-                {
-                    auto size = binsize[bin];
-                    byte* p = pool.baseAddr + pn * PAGESIZE;
-                    byte* ptop = p + PAGESIZE;
-                    size_t biti = pn * (PAGESIZE/16);
-                    size_t bitstride = size / 16;
 
-    version(none) // BUG: doesn't work because freebits() must also be cleared
+/**
+ * remove p from list of roots
+ */
+void removeRoot(void *p)
+{
+    if (!p)
     {
-                    // If free'd entire page
-                    if (bbase[0] == 0 && bbase[1] == 0 && bbase[2] == 0 && bbase[3] == 0 &&
-                        bbase[4] == 0 && bbase[5] == 0 && bbase[6] == 0 && bbase[7] == 0)
-                    {
-                        for (; p < ptop; p += size, biti += bitstride)
-                        {
-                            if (pool.finals.nbits && pool.finals.testClear(biti)) {
-                                if (opts.options.sentinel)
-                                    rt_finalize(cast(List *)sentinel_add(p), false/*noStack > 0*/);
-                                else
-                                    rt_finalize(cast(List *)p, false/*noStack > 0*/);
-                            }
-                            gcx.clrBits(pool, biti, BlkAttr.ALL_BITS);
+        return;
+    }
 
-                            List *list = cast(List *)p;
+    bool r;
+    if (!thread_needLock())
+    {
+        r = gc.roots.remove(p);
+    }
+    else synchronized (gc.lock)
+    {
+        r = gc.roots.remove(p);
+    }
+    assert (r);
+}
 
-                            if (opts.options.mem_stomp)
-                                memset(p, 0xF3, size);
-                        }
-                        pool.pagetable[pn] = B_FREE;
-                        freed += PAGESIZE;
-                        continue;
-                    }
+
+/**
+ * add range to scan for roots
+ */
+void addRange(void *p, size_t sz)
+{
+    if (!p || !sz)
+    {
+        return;
     }
-                    for (; p < ptop; p += size, biti += bitstride)
-                    {
-                        if (!pool.mark.test(biti))
-                        {
-                            if (opts.options.sentinel)
-                                sentinel_Invariant(sentinel_add(p));
-
-                            pool.freebits.set(biti);
-                            if (pool.finals.nbits && pool.finals.testClear(biti)) {
-                                if (opts.options.sentinel)
-                                    rt_finalize(cast(List *)sentinel_add(p), false/*noStack > 0*/);
-                                else
-                                    rt_finalize(cast(List *)p, false/*noStack > 0*/);
-                            }
-                            clrBits(pool, biti, BlkAttr.ALL_BITS);
 
-                            List *list = cast(List *)p;
+    if (!thread_needLock())
+    {
+        if (gc.ranges.append(Range(p, p+sz)) is null)
+            onOutOfMemoryError();
+    }
+    else synchronized (gc.lock)
+    {
+        if (gc.ranges.append(Range(p, p+sz)) is null)
+            onOutOfMemoryError();
+    }
+}
 
-                            if (opts.options.mem_stomp)
-                                memset(p, 0xF3, size);
 
-                            freed += size;
-                        }
-                    }
-                }
-                else if (bin == B_PAGE)
-                {
-                    size_t biti = pn * (PAGESIZE / 16);
-                    if (!pool.mark.test(biti))
-                    {
-                        byte *p = pool.baseAddr + pn * PAGESIZE;
-                        if (opts.options.sentinel)
-                            sentinel_Invariant(sentinel_add(p));
-                        if (pool.finals.nbits && pool.finals.testClear(biti)) {
-                            if (opts.options.sentinel)
-                                rt_finalize(sentinel_add(p), false/*noStack > 0*/);
-                            else
-                                rt_finalize(p, false/*noStack > 0*/);
-                        }
-                        clrBits(pool, biti, BlkAttr.ALL_BITS);
+/**
+ * remove range
+ */
+void removeRange(void *p)
+{
+    if (!p)
+    {
+        return;
+    }
 
-                        debug(COLLECT_PRINTF) printf("\tcollecting big %x\n", p);
-                        pool.pagetable[pn] = B_FREE;
-                        freedpages++;
-                        if (opts.options.mem_stomp)
-                            memset(p, 0xF3, PAGESIZE);
-                        while (pn + 1 < pool.npages && pool.pagetable[pn + 1] == B_PAGEPLUS)
-                        {
-                            pn++;
-                            pool.pagetable[pn] = B_FREE;
-                            freedpages++;
+    bool r;
+    if (!thread_needLock())
+    {
+        r = gc.ranges.remove(Range(p, null));
+    }
+    else synchronized (gc.lock)
+    {
+        r = gc.ranges.remove(Range(p, null));
+    }
+    assert (r);
+}
 
-                            if (opts.options.mem_stomp)
-                            {
-                                p += PAGESIZE;
-                                memset(p, 0xF3, PAGESIZE);
-                            }
-                        }
-                    }
-                }
-            }
-        }
 
-        // Zero buckets
-        bucket[] = null;
+/**
+ * do full garbage collection
+ */
+void fullCollect()
+{
 
-        // Free complete pages, rebuild free list
-        debug(COLLECT_PRINTF) printf("\tfree complete pages\n");
-        size_t recoveredpages = 0;
-        for (n = 0; n < pools.length; n++)
-        {
-            pool = pools[n];
-            for (size_t pn = 0; pn < pool.npages; pn++)
-            {
-                Bins   bin = cast(Bins)pool.pagetable[pn];
-                size_t biti;
-                size_t u;
+    if (!thread_needLock())
+    {
+        fullcollectshell();
+    }
+    else synchronized (gc.lock)
+    {
+        fullcollectshell();
+    }
 
-                if (bin < B_PAGE)
-                {
-                    size_t size = binsize[bin];
-                    size_t bitstride = size / 16;
-                    size_t bitbase = pn * (PAGESIZE / 16);
-                    size_t bittop = bitbase + (PAGESIZE / 16);
-                    byte*  p;
-
-                    biti = bitbase;
-                    for (biti = bitbase; biti < bittop; biti += bitstride)
-                    {
-                        if (!pool.freebits.test(biti))
-                            goto Lnotfree;
-                    }
-                    pool.pagetable[pn] = B_FREE;
-                    recoveredpages++;
-                    continue;
+    version (none)
+    {
+        GCStats stats;
+        getStats(stats);
+    }
 
-                 Lnotfree:
-                    p = pool.baseAddr + pn * PAGESIZE;
-                    for (u = 0; u < PAGESIZE; u += size)
-                    {
-                        biti = bitbase + u / 16;
-                        if (pool.freebits.test(biti))
-                        {
-                            List *list = cast(List *)(p + u);
-                            if (list.next != bucket[bin])       // avoid unnecessary writes
-                                list.next = bucket[bin];
-                            bucket[bin] = list;
-                        }
-                    }
-                }
-            }
-        }
+}
 
-        debug(COLLECT_PRINTF) printf("recovered pages = %d\n", recoveredpages);
-        debug(COLLECT_PRINTF) printf("\tfree'd %u bytes, %u pages from %u pools\n", freed, freedpages, pools.length);
 
-        return freedpages + recoveredpages;
+/**
+ * do full garbage collection ignoring roots
+ */
+void fullCollectNoStack()
+{
+    if (!thread_needLock())
+    {
+        gc.no_stack++;
+        fullcollectshell();
+        gc.no_stack--;
+    }
+    else synchronized (gc.lock)
+    {
+        gc.no_stack++;
+        fullcollectshell();
+        gc.no_stack--;
     }
+}
 
 
-    /**
-     *
-     */
-    uint getBits(Pool* pool, size_t biti)
-    in
+/**
+ * minimize free space usage
+ */
+void minimize()
+{
+    if (!thread_needLock())
     {
-        assert( pool );
+        minimizeNoSync();
     }
-    body
+    else synchronized (gc.lock)
     {
-        uint bits;
-
-        if (pool.finals.nbits &&
-            pool.finals.test(biti))
-            bits |= BlkAttr.FINALIZE;
-        if (pool.noscan.test(biti))
-            bits |= BlkAttr.NO_SCAN;
-//        if (pool.nomove.nbits &&
-//            pool.nomove.test(biti))
-//            bits |= BlkAttr.NO_MOVE;
-        return bits;
+        minimizeNoSync();
     }
+}
 
 
-    /**
-     *
-     */
-    void setBits(Pool* pool, size_t biti, uint mask)
-    in
+/**
+ * Retrieve statistics about garbage collection.
+ * Useful for debugging and tuning.
+ */
+void getStats(out GCStats stats)
+{
+    if (!thread_needLock())
     {
-        assert( pool );
+        getStatsNoSync(stats);
     }
-    body
+    else synchronized (gc.lock)
     {
-        if (mask & BlkAttr.FINALIZE)
-        {
-            if (!pool.finals.nbits)
-                pool.finals.alloc(pool.mark.nbits);
-            pool.finals.set(biti);
-        }
-        if (mask & BlkAttr.NO_SCAN)
+        getStatsNoSync(stats);
+    }
+}
+
+
+//
+//
+//
+private void getStatsNoSync(out GCStats stats)
+{
+    size_t psize = 0;
+    size_t usize = 0;
+    size_t flsize = 0;
+
+    size_t n;
+    size_t bsize = 0;
+
+    memset(&stats, 0, GCStats.sizeof);
+
+    for (n = 0; n < gc.pools.length; n++)
+    {
+        Pool* pool = gc.pools[n];
+        psize += pool.npages * PAGESIZE;
+        for (size_t j = 0; j < pool.npages; j++)
         {
-            pool.noscan.set(biti);
+            Bins bin = cast(Bins)pool.pagetable[j];
+            if (bin == B_FREE)
+                stats.freeblocks++;
+            else if (bin == B_PAGE)
+                stats.pageblocks++;
+            else if (bin < B_PAGE)
+                bsize += PAGESIZE;
         }
-//        if (mask & BlkAttr.NO_MOVE)
-//        {
-//            if (!pool.nomove.nbits)
-//                pool.nomove.alloc(pool.mark.nbits);
-//            pool.nomove.set(biti);
-//        }
     }
 
+    for (n = 0; n < B_PAGE; n++)
+    {
+        for (List *list = gc.free_list[n]; list; list = list.next)
+            flsize += binsize[n];
+    }
+
+    usize = bsize - flsize;
 
-    /**
-     *
-     */
-    void clrBits(Pool* pool, size_t biti, uint mask)
-    in
+    stats.poolsize = psize;
+    stats.usedsize = bsize - flsize;
+    stats.freelistsize = flsize;
+}
+
+/******************* weak-reference support *********************/
+
+// call locked if necessary
+private T locked(T)(in T delegate() code)
+{
+    if (thread_needLock)
+        synchronized(gc.lock) return code();
+    else
+       return code();
+}
+
+private struct WeakPointer
+{
+    Object reference;
+
+    void ondestroy(Object r)
     {
-        assert( pool );
+        assert(r is reference);
+        // lock for memory consistency (parallel readers)
+        // also ensures that weakpointerDestroy can be called while another
+        // thread is freeing the reference with "delete"
+        locked!(void)({ reference = null; });
+    }
+}
+
+/**
+ * Create a weak pointer to the given object.
+ * Returns a pointer to an opaque struct allocated in C memory.
+ */
+void* weakpointerCreate( Object r )
+{
+    if (r)
+    {
+        // must be allocated in C memory
+        // 1. to hide the reference from the GC
+        // 2. the GC doesn't scan delegates added by rt_attachDisposeEvent
+        //    for references
+        auto wp = cast(WeakPointer*)(cstdlib.malloc(WeakPointer.sizeof));
+        if (!wp)
+            onOutOfMemoryError();
+        wp.reference = r;
+        rt_attachDisposeEvent(r, &wp.ondestroy);
+        return wp;
     }
-    body
+    return null;
+}
+
+/**
+ * Destroy a weak pointer returned by weakpointerCreate().
+ * If null is passed, nothing happens.
+ */
+void weakpointerDestroy( void* p )
+{
+    if (p)
     {
-        if (mask & BlkAttr.FINALIZE && pool.finals.nbits)
-            pool.finals.clear(biti);
-        if (mask & BlkAttr.NO_SCAN)
-            pool.noscan.clear(biti);
-//        if (mask & BlkAttr.NO_MOVE && pool.nomove.nbits)
-//            pool.nomove.clear(biti);
+        auto wp = cast(WeakPointer*)p;
+        // must be extra careful about the GC or parallel threads
+        // finalizing the reference at the same time
+        locked!(void)({
+               if (wp.reference)
+                   rt_detachDisposeEvent(wp.reference, &wp.ondestroy);
+              });
+        cstdlib.free(wp);
     }
+}
 
+/**
+ * Query a weak pointer and return either the object passed to
+ * weakpointerCreate, or null if it was free'd in the meantime.
+ * If null is passed, null is returned.
+ */
+Object weakpointerGet( void* p )
+{
+    if (p)
+    {
+        // NOTE: could avoid the lock by using Fawzi style GC counters but
+        // that'd require core.sync.Atomic and lots of care about memory
+        // consistency it's an optional optimization see
+        // http://dsource.org/projects/tango/browser/trunk/user/tango/core/Lifetime.d?rev=5100#L158
+        return locked!(Object)({
+              return (cast(WeakPointer*)p).reference;
+              });
+        }
 }
 
 
@@ -2490,7 +2542,10 @@ struct Pool
     }
 
 
-    void Invariant() { }
+    bool Invariant()
+    {
+        return true;
+    }
 
 
     invariant
@@ -2602,4 +2657,222 @@ void *sentinel_sub(void *p)
 }
 
 
+
+/* ============================ C Public Interface ======================== */
+
+
+private int _termCleanupLevel=1;
+
+/// sets the cleanup level done by gc
+/// (0: none, 1: fullCollect, 2: fullCollectNoStack (might crash daemonThreads))
+/// result !=0 if the value was invalid
+extern (C) int gc_setTermCleanupLevel(int cLevel){
+    if (cLevel<0 || cLevel>2) return cLevel;
+    _termCleanupLevel=cLevel;
+    return 0;
+}
+
+/// returns the cleanup level done by gc
+extern (C) int gc_getTermCleanupLevel(){
+    return _termCleanupLevel;
+}
+
+version (DigitalMars) version(OSX) {
+    extern(C) void _d_osx_image_init();
+}
+
+extern (C) void thread_init();
+
+extern (C) void gc_init()
+{
+    scope (exit) assert (Invariant());
+    gc = cast(GC*) cstdlib.calloc(1, GC.sizeof);
+    *gc = GC.init;
+    initialize();
+    version (DigitalMars) version(OSX) {
+        _d_osx_image_init();
+    }
+    // NOTE: The GC must initialize the thread library
+    //       before its first collection.
+    thread_init();
+}
+
+extern (C) void gc_term()
+{
+    assert (Invariant());
+    if (_termCleanupLevel<1) {
+        // no cleanup
+    } else if (_termCleanupLevel==2){
+        // a more complete cleanup
+        // NOTE: There may be daemons threads still running when this routine is
+        //       called.  If so, cleaning memory out from under then is a good
+        //       way to make them crash horribly.
+        //       Often this probably doesn't matter much since the app is
+        //       supposed to be shutting down anyway, but for example tests might
+        //       crash (and be considerd failed even if the test was ok).
+        //       thus this is not the default and should be enabled by
+        //       I'm disabling cleanup for now until I can think about it some
+        //       more.
+        //
+        fullCollectNoStack(); // not really a 'collect all' -- still scans
+                                  // static data area, roots, and ranges.
+    } else {
+        // default (safe) clenup
+        fullCollect();
+    }
+}
+
+extern (C) void gc_enable()
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    enable();
+}
+
+extern (C) void gc_disable()
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    disable();
+}
+
+extern (C) void gc_collect()
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    fullCollect();
+}
+
+
+extern (C) void gc_minimize()
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    minimize();
+}
+
+extern (C) uint gc_getAttr( void* p )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    return getAttr(p);
+}
+
+extern (C) uint gc_setAttr( void* p, uint a )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    return setAttr(p, a);
+}
+
+extern (C) uint gc_clrAttr( void* p, uint a )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    return clrAttr(p, a);
+}
+
+extern (C) void* gc_malloc(size_t sz, uint attrs = 0,
+        PointerMap ptrmap = PointerMap.init)
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    return malloc(sz, attrs, ptrmap);
+}
+
+extern (C) void* gc_calloc(size_t sz, uint attrs = 0,
+        PointerMap ptrmap = PointerMap.init)
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    return calloc(sz, attrs, ptrmap);
+}
+
+extern (C) void* gc_realloc(void* p, size_t sz, uint attrs = 0,
+        PointerMap ptrmap = PointerMap.init)
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    return realloc(p, sz, attrs, ptrmap);
+}
+
+extern (C) size_t gc_extend( void* p, size_t mx, size_t sz )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    return extend(p, mx, sz);
+}
+
+extern (C) size_t gc_reserve( size_t sz )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    return reserve(sz);
+}
+
+extern (C) void gc_free( void* p )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    free(p);
+}
+
+extern (C) void* gc_addrOf( void* p )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    return addrOf(p);
+}
+
+extern (C) size_t gc_sizeOf( void* p )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    return sizeOf(p);
+}
+
+extern (C) BlkInfo gc_query( void* p )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    return query(p);
+}
+
+// NOTE: This routine is experimental.  The stats or function name may change
+//       before it is made officially available.
+extern (C) GCStats gc_stats()
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    GCStats stats = void;
+    getStats(stats);
+    return stats;
+}
+
+extern (C) void gc_addRoot( void* p )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    addRoot(p);
+}
+
+extern (C) void gc_addRange( void* p, size_t sz )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    addRange(p, sz);
+}
+
+extern (C) void gc_removeRoot( void *p )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    removeRoot(p);
+}
+
+extern (C) void gc_removeRange( void *p )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    removeRange(p);
+}
+
+extern (C) void* gc_weakpointerCreate( Object r )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    return weakpointerCreate(r);
+}
+
+extern (C) void gc_weakpointerDestroy( void* wp )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    weakpointerDestroy(wp);
+}
+
+extern (C) Object gc_weakpointerGet( void* wp )
+{
+    assert (Invariant()); scope (exit) assert (Invariant());
+    return weakpointerGet(wp);
+}
+
+
 // vim: set et sw=4 sts=4 :