Avoid output duplication because of FILE* buffers

[software/dgc/cdgc.git] / rt / gc / cdgc / gc.d

diff --git a/rt/gc/cdgc/gc.d b/rt/gc/cdgc/gc.d
index 267b42f04ebe1adc066af955a1f43a1e670a29b1..96c6a86360d24277f0456ea9038f17580f90864d 100644 (file)
--- a/rt/gc/cdgc/gc.d
+++ b/rt/gc/cdgc/gc.d
@@ -45,11 +45,12 @@ version = STACKGROWSDOWN;       // growing the stack means subtracting from the
 import rt.gc.cdgc.bits: GCBits;
 import rt.gc.cdgc.stats: GCStats, Stats;
 import dynarray = rt.gc.cdgc.dynarray;
 import rt.gc.cdgc.bits: GCBits;
 import rt.gc.cdgc.stats: GCStats, Stats;
 import dynarray = rt.gc.cdgc.dynarray;

-import alloc = rt.gc.cdgc.alloc;
+import os = rt.gc.cdgc.os;

 import opts = rt.gc.cdgc.opts;
 
 import cstdlib = tango.stdc.stdlib;
 import cstring = tango.stdc.string;
 import opts = rt.gc.cdgc.opts;
 
 import cstdlib = tango.stdc.stdlib;
 import cstring = tango.stdc.string;

+import cstdio = tango.stdc.stdio;

 
 /*
  * This is a small optimization that proved it's usefulness. For small chunks
 
 /*
  * This is a small optimization that proved it's usefulness. For small chunks


@@ -155,7 +156,8 @@ alias ubyte Bins;
 
 struct List
 {
 
 struct List
 {

-    List *next;
+    List* next;
+    Pool* pool;

 }
 
 
 }
 
 


@@ -210,7 +212,7 @@ struct GC
 
     dynarray.DynArray!(void*) roots;
     dynarray.DynArray!(Range) ranges;
 
     dynarray.DynArray!(void*) roots;
     dynarray.DynArray!(Range) ranges;

-    dynarray.DynArray!(Pool) pools;
+    dynarray.DynArray!(Pool*) pools;

 
     Stats stats;
 }
 
     Stats stats;
 }


@@ -236,7 +238,7 @@ bool Invariant()
             if (i == 0)
                 assert(gc.min_addr == pool.baseAddr);
             if (i + 1 < gc.pools.length)
             if (i == 0)
                 assert(gc.min_addr == pool.baseAddr);
             if (i + 1 < gc.pools.length)

-                assert(*pool < gc.pools[i + 1]);
+                assert(*pool < *gc.pools[i + 1]);

             else if (i + 1 == gc.pools.length)
                 assert(gc.max_addr == pool.topAddr);
         }
             else if (i + 1 == gc.pools.length)
                 assert(gc.max_addr == pool.topAddr);
         }


@@ -250,10 +252,14 @@ bool Invariant()
             assert(gc.ranges[i].pbot <= 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)
-            {
+        for (size_t i = 0; i < B_PAGE; i++) {
+            for (List *list = gc.free_list[i]; list; list = list.next) {
+                assert (list.pool !is null);
+                auto p = cast(byte*) list;
+                assert (p >= list.pool.baseAddr);
+                assert (p < list.pool.topAddr);

             }
             }

+        }

     }
     return true;
 }
     }
     return true;
 }


@@ -264,26 +270,28 @@ bool Invariant()
  * Return null if not in a Pool.
  * Assume pools is sorted.
  */
  * Return null if not in a Pool.
  * Assume pools is sorted.
  */

-Pool *findPool(void *p)
+Pool* findPool(void* p)

 {
 {

-    if (p >= gc.min_addr && p < gc.max_addr)
-    {
-        if (gc.pools.length == 1)
-        {
-            return gc.pools[0];
-        }
-
-        for (size_t i = 0; i < gc.pools.length; i++)
-        {
-            Pool* pool = gc.pools[i];
-            if (p < pool.topAddr)
-            {
-                if (pool.baseAddr <= p)
-                    return pool;
-                break;
-            }
-        }
+    if (p < gc.min_addr || p >= gc.max_addr)
+        return null;
+    if (gc.pools.length == 0)
+        return null;
+    if (gc.pools.length == 1)
+        return gc.pools[0];
+    /// The pooltable[] is sorted by address, so do a binary search
+    size_t low = 0;
+    size_t high = gc.pools.length - 1;
+    while (low <= high) {
+        size_t mid = (low + high) / 2;
+        auto pool = gc.pools[mid];
+        if (p < pool.baseAddr)
+            high = mid - 1;
+        else if (p >= pool.topAddr)
+            low = mid + 1;
+        else
+            return pool;

     }
     }

+    // Not found

     return null;
 }
 
     return null;
 }
 


@@ -300,10 +308,24 @@ BlkInfo getInfo(void* p)
         return BlkInfo.init;
     BlkInfo info;
     info.base = pool.findBase(p);
         return BlkInfo.init;
     BlkInfo info;
     info.base = pool.findBase(p);

+    if (info.base is null)
+        return BlkInfo.init;

     info.size = pool.findSize(info.base);
     info.attr = getAttr(pool, cast(size_t)(info.base - pool.baseAddr) / 16u);
     info.size = pool.findSize(info.base);
     info.attr = getAttr(pool, cast(size_t)(info.base - pool.baseAddr) / 16u);

-    if (!opts.options.conservative && !(info.attr & BlkAttr.NO_SCAN))
+    if (has_pointermap(info.attr)) {

         info.size -= size_t.sizeof; // PointerMap bitmask
         info.size -= size_t.sizeof; // PointerMap bitmask

+        // Points to the PointerMap bitmask pointer, not user data
+        if (p >= (info.base + info.size)) {
+            return BlkInfo.init;
+        }
+    }
+    if (opts.options.sentinel) {
+        info.base = sentinel_add(info.base);
+        // points to sentinel data, not user data
+        if (p < info.base || p >= sentinel_post(info.base))
+            return BlkInfo.init;
+        info.size -= SENTINEL_EXTRA;
+    }

     return info;
 }
 
     return info;
 }
 


@@ -311,7 +333,7 @@ BlkInfo getInfo(void* p)
 /**
  * Compute bin for size.
  */
 /**
  * Compute bin for size.
  */

-static Bins findBin(size_t size)
+Bins findBin(size_t size)

 {
     Bins bin;
     if (size <= 256)
 {
     Bins bin;
     if (size <= 256)


@@ -379,7 +401,7 @@ void minimize()
 {
     size_t n;
     size_t pn;
 {
     size_t n;
     size_t pn;

-    Pool*  pool;
+    Pool* pool;

 
     for (n = 0; n < gc.pools.length; n++)
     {
 
     for (n = 0; n < gc.pools.length; n++)
     {


@@ -392,6 +414,7 @@ void minimize()
         if (pn < pool.npages)
             continue;
         pool.Dtor();
         if (pn < pool.npages)
             continue;
         pool.Dtor();

+        cstdlib.free(pool);

         gc.pools.remove_at(n);
         n--;
     }
         gc.pools.remove_at(n);
         n--;
     }


@@ -404,9 +427,8 @@ void minimize()
  * Allocate a chunk of memory that is larger than a page.
  * Return null if out of memory.
  */
  * Allocate a chunk of memory that is larger than a page.
  * Return null if out of memory.
  */

-void *bigAlloc(size_t size)
+void* bigAlloc(size_t size, out Pool* pool)

 {
 {

-    Pool*  pool;

     size_t npages;
     size_t n;
     size_t pn;
     size_t npages;
     size_t n;
     size_t pn;


@@ -518,20 +540,24 @@ Pool *newPool(size_t npages)
             npages = n;
     }
 
             npages = n;
     }
 

-    Pool p;
-    p.initialize(npages);
-    if (!p.baseAddr)
+    auto pool = cast(Pool*) cstdlib.calloc(1, Pool.sizeof);
+    if (pool is null)
+        return null;
+    pool.initialize(npages);
+    if (!pool.baseAddr)

     {
     {

-        p.Dtor();
+        pool.Dtor();

         return null;
     }
 
         return null;
     }
 

-    Pool* pool = gc.pools.insert_sorted(p);
-    if (pool)
-    {
-        gc.min_addr = gc.pools[0].baseAddr;
-        gc.max_addr = gc.pools[gc.pools.length - 1].topAddr;
+    auto inserted_pool = *gc.pools.insert_sorted!("*a < *b")(pool);
+    if (inserted_pool is null) {
+        pool.Dtor();
+        return null;

     }
     }

+    assert (inserted_pool is pool);
+    gc.min_addr = gc.pools[0].baseAddr;
+    gc.max_addr = gc.pools[gc.pools.length - 1].topAddr;

     return pool;
 }
 
     return pool;
 }
 


@@ -563,33 +589,26 @@ int allocPage(Bins bin)
 
     // Convert page to free list
     size_t size = binsize[bin];
 
     // Convert page to free list
     size_t size = binsize[bin];

-    List **b = &gc.free_list[bin];
+    auto list_head = &gc.free_list[bin];

 
     p = pool.baseAddr + pn * PAGESIZE;
     ptop = p + PAGESIZE;
     for (; p < ptop; p += size)
     {
 
     p = pool.baseAddr + pn * PAGESIZE;
     ptop = p + PAGESIZE;
     for (; p < ptop; p += size)
     {

-        (cast(List *)p).next = *b;
-        *b = cast(List *)p;
+        List* l = cast(List *) p;
+        l.next = *list_head;
+        l.pool = pool;
+        *list_head = l;

     }
     return 1;
 }
 
 
 /**
     }
     return 1;
 }
 
 
 /**

- * Marks a range of memory using the conservative bit mask.  Used for
- * the stack, for the data segment, and additional memory ranges.
+ * Search a range of memory values and mark any pointers into the GC pool using
+ * type information (bitmask of pointer locations).

  */
  */

-void mark_conservative(void* pbot, void* ptop)
-{
-    mark(pbot, ptop, PointerMap.init.bits.ptr);
-}
-
-
-/**
- * Search a range of memory values and mark any pointers into the GC pool.
- */
-void mark(void *pbot, void *ptop, size_t* pm_bitmask)
+void mark_range(void *pbot, void *ptop, size_t* pm_bitmask)

 {
     // TODO: make our own assert because assert uses the GC
     assert (pbot <= ptop);
 {
     // TODO: make our own assert because assert uses the GC
     assert (pbot <= ptop);


@@ -599,16 +618,18 @@ void mark(void *pbot, void *ptop, size_t* pm_bitmask)
     void **p1 = cast(void **)pbot;
     void **p2 = cast(void **)ptop;
     size_t pcache = 0;
     void **p1 = cast(void **)pbot;
     void **p2 = cast(void **)ptop;
     size_t pcache = 0;

-    uint changes = 0;
+    bool changes = false;

 
     size_t type_size = pm_bitmask[0];
     size_t* pm_bits = pm_bitmask + 1;
 
     size_t type_size = pm_bitmask[0];
     size_t* pm_bits = pm_bitmask + 1;

+    bool has_type_info = type_size != 1 || pm_bits[0] != 1 || pm_bits[1] != 0;

 
     //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
 
     //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))))
+            if (has_type_info &&
+                    !(pm_bits[n / BITS_PER_WORD] & (1 << (n % BITS_PER_WORD))))

                 continue;
 
             void* p = *(p1 + n);
                 continue;
 
             void* p = *(p1 + n);


@@ -623,13 +644,17 @@ void mark(void *pbot, void *ptop, size_t* pm_bitmask)
             if (pool)
             {
                 size_t offset = cast(size_t)(p - pool.baseAddr);
             if (pool)
             {
                 size_t offset = cast(size_t)(p - pool.baseAddr);

-                size_t bit_i;
+                size_t bit_i = void;

                 size_t pn = offset / PAGESIZE;
                 Bins   bin = cast(Bins)pool.pagetable[pn];
 
                 size_t pn = offset / PAGESIZE;
                 Bins   bin = cast(Bins)pool.pagetable[pn];
 

+                // Cache B_PAGE, B_PAGEPLUS and B_FREE lookups
+                if (bin >= B_PAGE)
+                    pcache = cast(size_t)p & ~(PAGESIZE-1);
+

                 // Adjust bit to be at start of allocated memory block
                 if (bin <= B_PAGE)
                 // Adjust bit to be at start of allocated memory block
                 if (bin <= B_PAGE)

-                    bit_i = (offset & notbinsize[bin]) >> 4;
+                    bit_i = (offset & notbinsize[bin]) / 16;

                 else if (bin == B_PAGEPLUS)
                 {
                     do
                 else if (bin == B_PAGEPLUS)
                 {
                     do


@@ -639,14 +664,8 @@ void mark(void *pbot, void *ptop, size_t* pm_bitmask)
                     while (cast(Bins)pool.pagetable[pn] == B_PAGEPLUS);
                     bit_i = pn * (PAGESIZE / 16);
                 }
                     while (cast(Bins)pool.pagetable[pn] == B_PAGEPLUS);
                     bit_i = pn * (PAGESIZE / 16);
                 }

-                else
-                {
-                    // Don't mark bits in B_FREE pages
+                else // Don't mark bits in B_FREE pages

                     continue;
                     continue;

-                }
-
-                if (bin >= B_PAGE) // Cache B_PAGE and B_PAGEPLUS lookups
-                    pcache = cast(size_t)p & ~(PAGESIZE-1);

 
                 if (!pool.mark.test(bit_i))
                 {
 
                 if (!pool.mark.test(bit_i))
                 {


@@ -654,7 +673,7 @@ void mark(void *pbot, void *ptop, size_t* pm_bitmask)
                     if (!pool.noscan.test(bit_i))
                     {
                         pool.scan.set(bit_i);
                     if (!pool.noscan.test(bit_i))
                     {
                         pool.scan.set(bit_i);

-                        changes = 1;
+                        changes = true;

                     }
                 }
             }
                     }
                 }
             }


@@ -760,79 +779,120 @@ size_t fullcollectshell()
  */
 size_t fullcollect(void *stackTop)
 {
  */
 size_t fullcollect(void *stackTop)
 {

-    size_t n;
-    Pool*  pool;
-

     debug(COLLECT_PRINTF) printf("Gcx.fullcollect()\n");
 
     debug(COLLECT_PRINTF) printf("Gcx.fullcollect()\n");
 

+    // we always need to stop the world to make threads save the CPU registers
+    // in the stack and prepare themselves for thread_scanAll()

     thread_suspendAll();
     gc.stats.world_stopped();
 
     thread_suspendAll();
     gc.stats.world_stopped();
 

-    gc.p_cache = null;
-    gc.size_cache = 0;
+    if (opts.options.fork) {
+        cstdio.fflush(null); // avoid duplicated FILE* output
+        os.pid_t child_pid = os.fork();
+        assert (child_pid != -1); // don't accept errors in non-release mode
+        switch (child_pid) {
+        case -1: // if fork() fails, fallback to stop-the-world
+            opts.options.fork = false;
+            break;
+        case 0: // child process (i.e. the collectors mark phase)
+            mark(stackTop);
+            cstdlib.exit(0);
+            break; // bogus, will never reach here
+        default: // parent process (i.e. the mutator)
+            // start the world again and wait for the mark phase to finish
+            thread_resumeAll();
+            gc.stats.world_started();
+            int status = void;
+            os.pid_t wait_pid = os.waitpid(child_pid, &status, 0);
+            assert (wait_pid == child_pid);
+            return sweep();
+        }
+
+    }
+
+    // if we reach here, we are using the standard stop-the-world collection
+    mark(stackTop);
+    thread_resumeAll();
+    gc.stats.world_started();
+
+    return sweep();
+}
+
+
+/**
+ *
+ */
+void mark(void *stackTop)
+{
+    debug(COLLECT_PRINTF) printf("\tmark()\n");

 
     gc.any_changes = false;
 
     gc.any_changes = false;

-    for (n = 0; n < gc.pools.length; n++)
+    for (size_t n = 0; n < gc.pools.length; n++)

     {
     {

-        pool = gc.pools[n];
+        Pool* pool = gc.pools[n];

         pool.mark.zero();
         pool.scan.zero();
         pool.freebits.zero();
     }
 
     // Mark each free entry, so it doesn't get scanned
         pool.mark.zero();
         pool.scan.zero();
         pool.freebits.zero();
     }
 
     // Mark each free entry, so it doesn't get scanned

-    for (n = 0; n < B_PAGE; n++)
+    for (size_t n = 0; n < B_PAGE; n++)

     {
         for (List *list = gc.free_list[n]; list; list = list.next)
         {
     {
         for (List *list = gc.free_list[n]; list; list = list.next)
         {

-            pool = findPool(list);
-            assert(pool);
-            pool.freebits.set(cast(size_t)(cast(byte*)list - pool.baseAddr) / 16);
+            Pool* pool = list.pool;
+            auto ptr = cast(byte*) list;
+            assert (pool);
+            assert (pool.baseAddr <= ptr);
+            assert (ptr < pool.topAddr);
+            size_t bit_i = cast(size_t)(ptr - pool.baseAddr) / 16;
+            pool.freebits.set(bit_i);

         }
     }
 
         }
     }
 

-    for (n = 0; n < gc.pools.length; n++)
+    for (size_t n = 0; n < gc.pools.length; n++)

     {
     {

-        pool = gc.pools[n];
+        Pool* pool = gc.pools[n];

         pool.mark.copy(&pool.freebits);
     }
 
         pool.mark.copy(&pool.freebits);
     }
 

-    void mark_conservative_dg(void* pbot, void* ptop)
+    /// Marks a range of memory in conservative mode.
+    void mark_conservative_range(void* pbot, void* ptop)

     {
     {

-        mark_conservative(pbot, ptop);
+        mark_range(pbot, ptop, PointerMap.init.bits.ptr);

     }
 
     }
 

-    rt_scanStaticData(&mark_conservative_dg);
+    rt_scanStaticData(&mark_conservative_range);

 
     if (!gc.no_stack)
     {
         // Scan stacks and registers for each paused thread
 
     if (!gc.no_stack)
     {
         // Scan stacks and registers for each paused thread

-        thread_scanAll(&mark_conservative_dg, stackTop);
+        thread_scanAll(&mark_conservative_range, stackTop);

     }
 
     // Scan roots
     debug(COLLECT_PRINTF) printf("scan roots[]\n");
     }
 
     // Scan roots
     debug(COLLECT_PRINTF) printf("scan roots[]\n");

-    mark_conservative(gc.roots.ptr, gc.roots.ptr + gc.roots.length);
+    mark_conservative_range(gc.roots.ptr, gc.roots.ptr + gc.roots.length);

 
     // Scan ranges
     debug(COLLECT_PRINTF) printf("scan ranges[]\n");
 
     // Scan ranges
     debug(COLLECT_PRINTF) printf("scan ranges[]\n");

-    for (n = 0; n < gc.ranges.length; n++)
+    for (size_t n = 0; n < gc.ranges.length; n++)

     {
         debug(COLLECT_PRINTF) printf("\t%x .. %x\n", gc.ranges[n].pbot, gc.ranges[n].ptop);
     {
         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);
+        mark_conservative_range(gc.ranges[n].pbot, gc.ranges[n].ptop);

     }
 
     debug(COLLECT_PRINTF) printf("\tscan heap\n");
     while (gc.any_changes)
     {
         gc.any_changes = false;
     }
 
     debug(COLLECT_PRINTF) printf("\tscan heap\n");
     while (gc.any_changes)
     {
         gc.any_changes = false;

-        for (n = 0; n < gc.pools.length; n++)
+        for (size_t n = 0; n < gc.pools.length; n++)

         {
             uint *bbase;
             uint *b;
             uint *btop;
 
         {
             uint *bbase;
             uint *b;
             uint *btop;
 

-            pool = gc.pools[n];
+            Pool* pool = gc.pools[n];

 
             bbase = pool.scan.base();
             btop = bbase + pool.scan.nwords;
 
             bbase = pool.scan.base();
             btop = bbase + pool.scan.nwords;


@@ -867,12 +927,12 @@ size_t fullcollect(void *stackTop)
                     bin = cast(Bins)pool.pagetable[pn];
                     if (bin < B_PAGE) {
                         if (opts.options.conservative)
                     bin = cast(Bins)pool.pagetable[pn];
                     if (bin < B_PAGE) {
                         if (opts.options.conservative)

-                            mark_conservative(o, o + binsize[bin]);
+                            mark_conservative_range(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;
                         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);
+                            mark_range(o, end_of_blk, pm_bitmask);

                         }
                     }
                     else if (bin == B_PAGE || bin == B_PAGEPLUS)
                         }
                     }
                     else if (bin == B_PAGE || bin == B_PAGEPLUS)


@@ -889,29 +949,36 @@ size_t fullcollect(void *stackTop)
 
                         size_t blk_size = u * PAGESIZE;
                         if (opts.options.conservative)
 
                         size_t blk_size = u * PAGESIZE;
                         if (opts.options.conservative)

-                            mark_conservative(o, o + blk_size);
+                            mark_conservative_range(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;
                         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);
+                            mark_range(o, end_of_blk, pm_bitmask);

                         }
                     }
                 }
             }
         }
     }
                         }
                     }
                 }
             }
         }
     }

+}

 
 

-    thread_resumeAll();
-    gc.stats.world_started();

 
 

+/**
+ *
+ */
+size_t sweep()
+{

     // Free up everything not marked
     // Free up everything not marked

-    debug(COLLECT_PRINTF) printf("\tfree'ing\n");
+    debug(COLLECT_PRINTF) printf("\tsweep\n");
+    gc.p_cache = null;
+    gc.size_cache = 0;

     size_t freedpages = 0;
     size_t freed = 0;
     size_t freedpages = 0;
     size_t freed = 0;

-    for (n = 0; n < gc.pools.length; n++)
+    for (size_t n = 0; n < gc.pools.length; n++)

     {
     {

-        pool = gc.pools[n];
+        Pool* pool = gc.pools[n];
+        pool.clear_cache();

         uint*  bbase = pool.mark.base();
         size_t pn;
         for (pn = 0; pn < pool.npages; pn++, bbase += PAGESIZE / (32 * 16))
         uint*  bbase = pool.mark.base();
         size_t pn;
         for (pn = 0; pn < pool.npages; pn++, bbase += PAGESIZE / (32 * 16))


@@ -935,16 +1002,14 @@ version(none) // BUG: doesn't work because freebits() must also be cleared
                 {
                     for (; p < ptop; p += size, bit_i += bit_stride)
                     {
                 {
                     for (; p < ptop; p += size, bit_i += bit_stride)
                     {

-                        if (pool.finals.nbits && pool.finals.testClear(bit_i)) {
+                        if (pool.finals.testClear(bit_i)) {

                             if (opts.options.sentinel)
                             if (opts.options.sentinel)

-                                rt_finalize(cast(List *)sentinel_add(p), false/*gc.no_stack > 0*/);
+                                rt_finalize(sentinel_add(p), false/*gc.no_stack > 0*/);

                             else
                             else

-                                rt_finalize(cast(List *)p, false/*gc.no_stack > 0*/);
+                                rt_finalize(p, false/*gc.no_stack > 0*/);

                         }
                         clrAttr(pool, bit_i, BlkAttr.ALL_BITS);
 
                         }
                         clrAttr(pool, bit_i, BlkAttr.ALL_BITS);
 

-                        List *list = cast(List *)p;
-

                         if (opts.options.mem_stomp)
                             memset(p, 0xF3, size);
                     }
                         if (opts.options.mem_stomp)
                             memset(p, 0xF3, size);
                     }


@@ -961,16 +1026,14 @@ version(none) // BUG: doesn't work because freebits() must also be cleared
                             sentinel_Invariant(sentinel_add(p));
 
                         pool.freebits.set(bit_i);
                             sentinel_Invariant(sentinel_add(p));
 
                         pool.freebits.set(bit_i);

-                        if (pool.finals.nbits && pool.finals.testClear(bit_i)) {
+                        if (pool.finals.testClear(bit_i)) {

                             if (opts.options.sentinel)
                             if (opts.options.sentinel)

-                                rt_finalize(cast(List *)sentinel_add(p), false/*gc.no_stack > 0*/);
+                                rt_finalize(sentinel_add(p), false/*gc.no_stack > 0*/);

                             else
                             else

-                                rt_finalize(cast(List *)p, false/*gc.no_stack > 0*/);
+                                rt_finalize(p, false/*gc.no_stack > 0*/);

                         }
                         clrAttr(pool, bit_i, BlkAttr.ALL_BITS);
 
                         }
                         clrAttr(pool, bit_i, BlkAttr.ALL_BITS);
 

-                        List *list = cast(List *)p;
-

                         if (opts.options.mem_stomp)
                             memset(p, 0xF3, size);
 
                         if (opts.options.mem_stomp)
                             memset(p, 0xF3, size);
 


@@ -986,7 +1049,7 @@ version(none) // BUG: doesn't work because freebits() must also be cleared
                     byte *p = pool.baseAddr + pn * PAGESIZE;
                     if (opts.options.sentinel)
                         sentinel_Invariant(sentinel_add(p));
                     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 (pool.finals.testClear(bit_i)) {

                         if (opts.options.sentinel)
                             rt_finalize(sentinel_add(p), false/*gc.no_stack > 0*/);
                         else
                         if (opts.options.sentinel)
                             rt_finalize(sentinel_add(p), false/*gc.no_stack > 0*/);
                         else


@@ -1022,9 +1085,9 @@ version(none) // BUG: doesn't work because freebits() must also be cleared
     // Free complete pages, rebuild free list
     debug(COLLECT_PRINTF) printf("\tfree complete pages\n");
     size_t recoveredpages = 0;
     // 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++)
+    for (size_t n = 0; n < gc.pools.length; n++)

     {
     {

-        pool = gc.pools[n];
+        Pool* pool = gc.pools[n];

         for (size_t pn = 0; pn < pool.npages; pn++)
         {
             Bins   bin = cast(Bins)pool.pagetable[pn];
         for (size_t pn = 0; pn < pool.npages; pn++)
         {
             Bins   bin = cast(Bins)pool.pagetable[pn];


@@ -1056,10 +1119,14 @@ version(none) // BUG: doesn't work because freebits() must also be cleared
                     bit_i = bit_base + u / 16;
                     if (pool.freebits.test(bit_i))
                     {
                     bit_i = bit_base + u / 16;
                     if (pool.freebits.test(bit_i))
                     {

-                        List *list = cast(List *)(p + u);
-                        // avoid unnecessary writes
+                        assert ((p+u) >= pool.baseAddr);
+                        assert ((p+u) < pool.topAddr);
+                        List* list = cast(List*) (p + u);
+                        // avoid unnecesary writes (it really saves time)

                         if (list.next != gc.free_list[bin])
                             list.next = gc.free_list[bin];
                         if (list.next != gc.free_list[bin])
                             list.next = gc.free_list[bin];

+                        if (list.pool != pool)
+                            list.pool = pool;

                         gc.free_list[bin] = list;
                     }
                 }
                         gc.free_list[bin] = list;
                     }
                 }


@@ -1085,14 +1152,11 @@ in
 body
 {
     uint attrs;
 body
 {
     uint attrs;

-
-    if (pool.finals.nbits &&
-        pool.finals.test(bit_i))
+    if (pool.finals.test(bit_i))

         attrs |= BlkAttr.FINALIZE;
     if (pool.noscan.test(bit_i))
         attrs |= BlkAttr.NO_SCAN;
         attrs |= BlkAttr.FINALIZE;
     if (pool.noscan.test(bit_i))
         attrs |= BlkAttr.NO_SCAN;

-//        if (pool.nomove.nbits &&
-//            pool.nomove.test(bit_i))
+//        if (pool.nomove.test(bit_i))

 //            attrs |= BlkAttr.NO_MOVE;
     return attrs;
 }
 //            attrs |= BlkAttr.NO_MOVE;
     return attrs;
 }


@@ -1110,8 +1174,6 @@ body
 {
     if (mask & BlkAttr.FINALIZE)
     {
 {
     if (mask & BlkAttr.FINALIZE)
     {

-        if (!pool.finals.nbits)
-            pool.finals.alloc(pool.mark.nbits);

         pool.finals.set(bit_i);
     }
     if (mask & BlkAttr.NO_SCAN)
         pool.finals.set(bit_i);
     }
     if (mask & BlkAttr.NO_SCAN)


@@ -1137,7 +1199,7 @@ in
 }
 body
 {
 }
 body
 {

-    if (mask & BlkAttr.FINALIZE && pool.finals.nbits)
+    if (mask & BlkAttr.FINALIZE)

         pool.finals.clear(bit_i);
     if (mask & BlkAttr.NO_SCAN)
         pool.noscan.clear(bit_i);
         pool.finals.clear(bit_i);
     if (mask & BlkAttr.NO_SCAN)
         pool.noscan.clear(bit_i);


@@ -1152,6 +1214,9 @@ void initialize()
     int dummy;
     gc.stack_bottom = cast(char*)&dummy;
     opts.parse(cstdlib.getenv("D_GC_OPTS"));
     int dummy;
     gc.stack_bottom = cast(char*)&dummy;
     opts.parse(cstdlib.getenv("D_GC_OPTS"));

+    // If we are going to fork, make sure we have the needed OS support
+    if (opts.options.fork)
+        opts.options.fork = os.HAVE_SHARED && os.HAVE_FORK;

     gc.lock = GCLock.classinfo;
     gc.inited = 1;
     setStackBottom(rt_stackBottom());
     gc.lock = GCLock.classinfo;
     gc.inited = 1;
     setStackBottom(rt_stackBottom());


@@ -1193,7 +1258,8 @@ private void *malloc(size_t size, uint attrs, size_t* pm_bitmask)
         lastbin = bin;
     }
 
         lastbin = bin;
     }
 

-    size_t capacity; // to figure out where to store the bitmask
+    Pool* pool = void;
+    size_t capacity = void; // to figure out where to store the bitmask

     if (bin < B_PAGE)
     {
         p = gc.free_list[bin];
     if (bin < B_PAGE)
     {
         p = gc.free_list[bin];


@@ -1220,6 +1286,7 @@ private void *malloc(size_t size, uint attrs, size_t* pm_bitmask)
             if (!gc.free_list[bin] && !allocPage(bin))
             {
                 newPool(1);         // allocate new pool to find a new page
             if (!gc.free_list[bin] && !allocPage(bin))
             {
                 newPool(1);         // allocate new pool to find a new page

+                // TODO: hint allocPage() to use the pool we just created

                 int result = allocPage(bin);
                 if (!result)
                     onOutOfMemoryError();
                 int result = allocPage(bin);
                 if (!result)
                     onOutOfMemoryError();


@@ -1229,7 +1296,11 @@ private void *malloc(size_t size, uint attrs, size_t* pm_bitmask)
         capacity = binsize[bin];
 
         // Return next item from free list
         capacity = binsize[bin];
 
         // Return next item from free list

-        gc.free_list[bin] = (cast(List*)p).next;
+        List* list = cast(List*) p;
+        assert ((cast(byte*)list) >= list.pool.baseAddr);
+        assert ((cast(byte*)list) < list.pool.topAddr);
+        gc.free_list[bin] = list.next;
+        pool = list.pool;

         if (!(attrs & BlkAttr.NO_SCAN))
             memset(p + size, 0, capacity - size);
         if (opts.options.mem_stomp)
         if (!(attrs & BlkAttr.NO_SCAN))
             memset(p + size, 0, capacity - size);
         if (opts.options.mem_stomp)


@@ -1237,9 +1308,10 @@ private void *malloc(size_t size, uint attrs, size_t* pm_bitmask)
     }
     else
     {
     }
     else
     {

-        p = bigAlloc(size);
+        p = bigAlloc(size, pool);

         if (!p)
             onOutOfMemoryError();
         if (!p)
             onOutOfMemoryError();

+        assert (pool !is null);

         // Round the size up to the number of pages needed to store it
         size_t npages = (size + PAGESIZE - 1) / PAGESIZE;
         capacity = npages * PAGESIZE;
         // Round the size up to the number of pages needed to store it
         size_t npages = (size + PAGESIZE - 1) / PAGESIZE;
         capacity = npages * PAGESIZE;


@@ -1260,12 +1332,8 @@ private void *malloc(size_t size, uint attrs, size_t* pm_bitmask)
     }
 
     if (attrs)
     }
 
     if (attrs)

-    {
-        Pool *pool = findPool(p);
-        assert(pool);
-

         setAttr(pool, cast(size_t)(p - pool.baseAddr) / 16, attrs);
         setAttr(pool, cast(size_t)(p - pool.baseAddr) / 16, attrs);

-    }
+

     return p;
 }
 
     return p;
 }
 


@@ -1362,10 +1430,13 @@ private void *realloc(void *p, size_t size, uint attrs,
                         memset(p + size - pm_bitmask_size, 0xF2,
                                 blk_size - size - pm_bitmask_size);
                     pool.freePages(pagenum + newsz, psz - newsz);
                         memset(p + size - pm_bitmask_size, 0xF2,
                                 blk_size - size - pm_bitmask_size);
                     pool.freePages(pagenum + newsz, psz - newsz);

+                    auto new_blk_size = (PAGESIZE * newsz);
+                    // update the size cache, assuming that is very likely the
+                    // size of this block will be queried in the near future
+                    pool.update_cache(p, new_blk_size);

                     if (has_pm) {
                     if (has_pm) {

-                        auto end_of_blk = cast(size_t**)(
-                                blk_base_addr + (PAGESIZE * newsz) -
-                                pm_bitmask_size);
+                        auto end_of_blk = cast(size_t**)(blk_base_addr +
+                                new_blk_size - pm_bitmask_size);

                         *end_of_blk = pm_bitmask;
                     }
                     return p;
                         *end_of_blk = pm_bitmask;
                     }
                     return p;


@@ -1383,10 +1454,14 @@ private void *realloc(void *p, size_t size, uint attrs,
                                         - pm_bitmask_size);
                             memset(pool.pagetable + pagenum +
                                     psz, B_PAGEPLUS, newsz - psz);
                                         - pm_bitmask_size);
                             memset(pool.pagetable + pagenum +
                                     psz, B_PAGEPLUS, newsz - psz);

+                            auto new_blk_size = (PAGESIZE * newsz);
+                            // update the size cache, assuming that is very
+                            // likely the size of this block will be queried in
+                            // the near future
+                            pool.update_cache(p, new_blk_size);

                             if (has_pm) {
                                 auto end_of_blk = cast(size_t**)(
                             if (has_pm) {
                                 auto end_of_blk = cast(size_t**)(

-                                        blk_base_addr +
-                                        (PAGESIZE * newsz) -
+                                        blk_base_addr + new_blk_size -

                                         pm_bitmask_size);
                                 *end_of_blk = pm_bitmask;
                             }
                                         pm_bitmask_size);
                                 *end_of_blk = pm_bitmask;
                             }


@@ -1495,6 +1570,9 @@ body
     memset(pool.pagetable + pagenum + psz, B_PAGEPLUS, sz);
     gc.p_cache = null;
     gc.size_cache = 0;
     memset(pool.pagetable + pagenum + psz, B_PAGEPLUS, sz);
     gc.p_cache = null;
     gc.size_cache = 0;

+    // update the size cache, assuming that is very likely the size of this
+    // block will be queried in the near future
+    pool.update_cache(p, new_size);

 
     if (has_pm) {
         new_size -= size_t.sizeof;
 
     if (has_pm) {
         new_size -= size_t.sizeof;


@@ -1540,16 +1618,19 @@ private void free(void *p)
         if (opts.options.mem_stomp)
             memset(p, 0xF2, npages * PAGESIZE);
         pool.freePages(pagenum, npages);
         if (opts.options.mem_stomp)
             memset(p, 0xF2, npages * PAGESIZE);
         pool.freePages(pagenum, npages);

+        // just in case we were caching this pointer
+        pool.clear_cache(p);

     }
     else
     {
         // Add to free list
     }
     else
     {
         // Add to free list

-        List *list = cast(List*)p;
+        List* list = cast(List*) p;

 
         if (opts.options.mem_stomp)
             memset(p, 0xF2, binsize[bin]);
 
         list.next = gc.free_list[bin];
 
         if (opts.options.mem_stomp)
             memset(p, 0xF2, binsize[bin]);
 
         list.next = gc.free_list[bin];

+        list.pool = pool;

         gc.free_list[bin] = list;
     }
 }
         gc.free_list[bin] = list;
     }
 }


@@ -1640,9 +1721,7 @@ private void checkNoSync(void *p)
             if (bin < B_PAGE)
             {
                 // Check that p is not on a free list
             if (bin < B_PAGE)
             {
                 // Check that p is not on a free list

-                List *list;
-
-                for (list = gc.free_list[bin]; list; list = list.next)
+                for (List* list = gc.free_list[bin]; list; list = list.next)

                 {
                     assert(cast(void*)list != p);
                 }
                 {
                     assert(cast(void*)list != p);
                 }


@@ -1708,7 +1787,7 @@ private GCStats getStats()
 
     for (n = 0; n < B_PAGE; n++)
     {
 
     for (n = 0; n < B_PAGE; n++)
     {

-        for (List *list = gc.free_list[n]; list; list = list.next)
+        for (List* list = gc.free_list[n]; list; list = list.next)

             flsize += binsize[n];
     }
 
             flsize += binsize[n];
     }
 


@@ -1815,12 +1894,29 @@ struct Pool
     size_t npages;
     ubyte* pagetable;
 
     size_t npages;
     ubyte* pagetable;
 

+    /// Cache for findSize()
+    size_t cached_size;
+    void* cached_ptr;
+
+    void clear_cache(void* ptr = null)
+    {
+        if (ptr is null || ptr is this.cached_ptr) {
+            this.cached_ptr = null;
+            this.cached_size = 0;
+        }
+    }
+
+    void update_cache(void* ptr, size_t size)
+    {
+        this.cached_ptr = ptr;
+        this.cached_size = size;
+    }

 
     void initialize(size_t npages)
     {
         size_t poolsize = npages * PAGESIZE;
         assert(poolsize >= POOLSIZE);
 
     void initialize(size_t npages)
     {
         size_t poolsize = npages * PAGESIZE;
         assert(poolsize >= POOLSIZE);

-        baseAddr = cast(byte *) alloc.os_mem_map(poolsize);
+        baseAddr = cast(byte *) os.alloc(poolsize);

 
         // Some of the code depends on page alignment of memory pools
         assert((cast(size_t)baseAddr & (PAGESIZE - 1)) == 0);
 
         // Some of the code depends on page alignment of memory pools
         assert((cast(size_t)baseAddr & (PAGESIZE - 1)) == 0);


@@ -1830,13 +1926,20 @@ struct Pool
             npages = 0;
             poolsize = 0;
         }
             npages = 0;
             poolsize = 0;
         }

-        //assert(baseAddr);

         topAddr = baseAddr + poolsize;
 
         topAddr = baseAddr + poolsize;
 

-        mark.alloc(cast(size_t)poolsize / 16);
-        scan.alloc(cast(size_t)poolsize / 16);
-        freebits.alloc(cast(size_t)poolsize / 16);
-        noscan.alloc(cast(size_t)poolsize / 16);
+        size_t nbits = cast(size_t)poolsize / 16;
+
+        // if the GC will run in parallel in a fork()ed process, we need to
+        // share the mark bits
+        os.Vis vis = os.Vis.PRIV;
+        if (opts.options.fork)
+            vis = os.Vis.SHARED;
+        mark.alloc(nbits, vis); // shared between mark and sweep
+        freebits.alloc(nbits, vis); // ditto
+        scan.alloc(nbits); // only used in the mark phase
+        finals.alloc(nbits); // mark phase *MUST* have a snapshot
+        noscan.alloc(nbits); // ditto

 
         pagetable = cast(ubyte*) cstdlib.malloc(npages);
         if (!pagetable)
 
         pagetable = cast(ubyte*) cstdlib.malloc(npages);
         if (!pagetable)


@@ -1855,7 +1958,7 @@ struct Pool
 
             if (npages)
             {
 
             if (npages)
             {

-                result = alloc.os_mem_unmap(baseAddr, npages * PAGESIZE);
+                result = os.dealloc(baseAddr, npages * PAGESIZE);

                 assert(result);
                 npages = 0;
             }
                 assert(result);
                 npages = 0;
             }


@@ -1867,9 +1970,12 @@ struct Pool
         if (pagetable)
             cstdlib.free(pagetable);
 
         if (pagetable)
             cstdlib.free(pagetable);
 

-        mark.Dtor();
+        os.Vis vis = os.Vis.PRIV;
+        if (opts.options.fork)
+            vis = os.Vis.SHARED;
+        mark.Dtor(vis);
+        freebits.Dtor(vis);

         scan.Dtor();
         scan.Dtor();

-        freebits.Dtor();

         finals.Dtor();
         noscan.Dtor();
     }
         finals.Dtor();
         noscan.Dtor();
     }


@@ -1971,10 +2077,15 @@ struct Pool
         Bins bin = cast(Bins)this.pagetable[pagenum];
         if (bin != B_PAGE)
             return binsize[bin];
         Bins bin = cast(Bins)this.pagetable[pagenum];
         if (bin != B_PAGE)
             return binsize[bin];

-        for (size_t i = pagenum + 1; i < this.npages; i++)
+        if (this.cached_ptr == p)
+            return this.cached_size;
+        size_t i = pagenum + 1;
+        for (; i < this.npages; i++)

             if (this.pagetable[i] != B_PAGEPLUS)
             if (this.pagetable[i] != B_PAGEPLUS)

-                return (i - pagenum) * PAGESIZE;
-        return (this.npages - pagenum) * PAGESIZE;
+                break;
+        this.cached_ptr = p;
+        this.cached_size = (i - pagenum) * PAGESIZE;
+        return this.cached_size;

     }
 
 
     }
 
 


@@ -2014,8 +2125,9 @@ void sentinel_init(void *p, size_t size)
 
 void sentinel_Invariant(void *p)
 {
 
 void sentinel_Invariant(void *p)
 {

-    assert(*sentinel_pre(p) == SENTINEL_PRE);
-    assert(*sentinel_post(p) == SENTINEL_POST);
+    if (*sentinel_pre(p) != SENTINEL_PRE ||
+            *sentinel_post(p) != SENTINEL_POST)
+        cstdlib.abort();

 }
 
 
 }