+}
+#endif
+
+#if defined(THREADED_RTS)
+static void
+gc_thread_entry (gc_thread *my_gct)
+{
+ gct = my_gct;
+ debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
+ gct->id = osThreadId();
+ gc_thread_mainloop();
+}
+#endif
+
+static void
+start_gc_threads (void)
+{
+#if defined(THREADED_RTS)
+ nat i;
+ OSThreadId id;
+ static rtsBool done = rtsFalse;
+
+ gc_running_threads = 0;
+ initMutex(&gc_running_mutex);
+
+ if (!done) {
+ // Start from 1: the main thread is 0
+ for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
+ createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
+ gc_threads[i]);
+ }
+ done = rtsTrue;
+ }
+#endif
+}
+
+static void
+wakeup_gc_threads (nat n_threads USED_IF_THREADS)
+{
+#if defined(THREADED_RTS)
+ nat i;
+ for (i=1; i < n_threads; i++) {
+ inc_running();
+ debugTrace(DEBUG_gc, "waking up gc thread %d", i);
+ do {
+ ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
+ if (gc_threads[i]->wakeup) {
+ RELEASE_LOCK(&gc_threads[i]->wake_mutex);
+ continue;
+ } else {
+ break;
+ }
+ } while (1);
+ gc_threads[i]->wakeup = rtsTrue;
+ signalCondition(&gc_threads[i]->wake_cond);
+ RELEASE_LOCK(&gc_threads[i]->wake_mutex);
+ }
+#endif
+}
+
+// After GC is complete, we must wait for all GC threads to enter the
+// standby state, otherwise they may still be executing inside
+// any_work(), and may even remain awake until the next GC starts.
+static void
+shutdown_gc_threads (nat n_threads USED_IF_THREADS)
+{
+#if defined(THREADED_RTS)
+ nat i;
+ rtsBool wakeup;
+ for (i=1; i < n_threads; i++) {
+ do {
+ ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
+ wakeup = gc_threads[i]->wakeup;
+ // wakeup is false while the thread is waiting
+ RELEASE_LOCK(&gc_threads[i]->wake_mutex);
+ } while (wakeup);
+ }
+#endif
+}
+
+/* ----------------------------------------------------------------------------
+ Initialise a generation that is to be collected
+ ------------------------------------------------------------------------- */
+
+static void
+init_collected_gen (nat g, nat n_threads)
+{
+ nat s, t, i;
+ step_workspace *ws;
+ step *stp;
+ bdescr *bd;
+
+ // Throw away the current mutable list. Invariant: the mutable
+ // list always has at least one block; this means we can avoid a
+ // check for NULL in recordMutable().
+ if (g != 0) {
+ freeChain(generations[g].mut_list);
+ generations[g].mut_list = allocBlock();
+ for (i = 0; i < n_capabilities; i++) {
+ freeChain(capabilities[i].mut_lists[g]);
+ capabilities[i].mut_lists[g] = allocBlock();
+ }
+ }
+
+ for (s = 0; s < generations[g].n_steps; s++) {
+
+ stp = &generations[g].steps[s];
+ ASSERT(stp->gen_no == g);
+
+ // we'll construct a new list of threads in this step
+ // during GC, throw away the current list.
+ stp->old_threads = stp->threads;
+ stp->threads = END_TSO_QUEUE;
+
+ // generation 0, step 0 doesn't need to-space
+ if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
+ continue;
+ }
+
+ // deprecate the existing blocks
+ stp->old_blocks = stp->blocks;
+ stp->n_old_blocks = stp->n_blocks;
+ stp->blocks = NULL;
+ stp->n_blocks = 0;
+ stp->n_words = 0;
+ stp->live_estimate = 0;
+
+ // we don't have any to-be-scavenged blocks yet
+ stp->todos = NULL;
+ stp->todos_last = NULL;
+ stp->n_todos = 0;
+
+ // initialise the large object queues.
+ stp->scavenged_large_objects = NULL;
+ stp->n_scavenged_large_blocks = 0;
+
+ // mark the small objects as from-space
+ for (bd = stp->old_blocks; bd; bd = bd->link) {
+ bd->flags &= ~BF_EVACUATED;
+ }
+
+ // mark the large objects as from-space
+ for (bd = stp->large_objects; bd; bd = bd->link) {
+ bd->flags &= ~BF_EVACUATED;
+ }
+
+ // for a compacted step, we need to allocate the bitmap
+ if (stp->mark) {
+ nat bitmap_size; // in bytes
+ bdescr *bitmap_bdescr;
+ StgWord *bitmap;
+
+ bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
+
+ if (bitmap_size > 0) {
+ bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
+ / BLOCK_SIZE);
+ stp->bitmap = bitmap_bdescr;
+ bitmap = bitmap_bdescr->start;
+
+ debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
+ bitmap_size, bitmap);
+
+ // don't forget to fill it with zeros!
+ memset(bitmap, 0, bitmap_size);
+
+ // For each block in this step, point to its bitmap from the
+ // block descriptor.
+ for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
+ bd->u.bitmap = bitmap;
+ bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
+
+ // Also at this point we set the BF_MARKED flag
+ // for this block. The invariant is that
+ // BF_MARKED is always unset, except during GC
+ // when it is set on those blocks which will be
+ // compacted.
+ if (!(bd->flags & BF_FRAGMENTED)) {
+ bd->flags |= BF_MARKED;
+ }
+ }
+ }
+ }
+ }
+
+ // For each GC thread, for each step, allocate a "todo" block to
+ // store evacuated objects to be scavenged, and a block to store
+ // evacuated objects that do not need to be scavenged.
+ for (t = 0; t < n_threads; t++) {
+ for (s = 0; s < generations[g].n_steps; s++) {
+
+ // we don't copy objects into g0s0, unless -G0
+ if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
+
+ ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
+
+ ws->todo_large_objects = NULL;
+
+ ws->part_list = NULL;
+ ws->n_part_blocks = 0;
+
+ // allocate the first to-space block; extra blocks will be
+ // chained on as necessary.
+ ws->todo_bd = NULL;
+ ws->buffer_todo_bd = NULL;
+ alloc_todo_block(ws,0);
+
+ ws->scavd_list = NULL;
+ ws->n_scavd_blocks = 0;
+ }
+ }
+}
+
+
+/* ----------------------------------------------------------------------------
+ Initialise a generation that is *not* to be collected
+ ------------------------------------------------------------------------- */
+
+static void
+init_uncollected_gen (nat g, nat threads)
+{
+ nat s, t, i;
+ step_workspace *ws;
+ step *stp;
+ bdescr *bd;
+
+ for (s = 0; s < generations[g].n_steps; s++) {
+ stp = &generations[g].steps[s];
+ stp->scavenged_large_objects = NULL;
+ stp->n_scavenged_large_blocks = 0;
+ }
+
+ for (s = 0; s < generations[g].n_steps; s++) {
+
+ stp = &generations[g].steps[s];
+
+ for (t = 0; t < threads; t++) {
+ ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
+
+ ws->buffer_todo_bd = NULL;
+ ws->todo_large_objects = NULL;
+
+ ws->part_list = NULL;
+ ws->n_part_blocks = 0;
+
+ ws->scavd_list = NULL;
+ ws->n_scavd_blocks = 0;
+
+ // If the block at the head of the list in this generation
+ // is less than 3/4 full, then use it as a todo block.
+ if (stp->blocks && isPartiallyFull(stp->blocks))
+ {
+ ws->todo_bd = stp->blocks;
+ ws->todo_free = ws->todo_bd->free;
+ ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
+ stp->blocks = stp->blocks->link;
+ stp->n_blocks -= 1;
+ stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
+ ws->todo_bd->link = NULL;
+ // we must scan from the current end point.
+ ws->todo_bd->u.scan = ws->todo_bd->free;
+ }
+ else
+ {
+ ws->todo_bd = NULL;
+ alloc_todo_block(ws,0);
+ }
+ }
+
+ // deal out any more partial blocks to the threads' part_lists
+ t = 0;
+ while (stp->blocks && isPartiallyFull(stp->blocks))
+ {
+ bd = stp->blocks;
+ stp->blocks = bd->link;
+ ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
+ bd->link = ws->part_list;
+ ws->part_list = bd;
+ ws->n_part_blocks += 1;
+ bd->u.scan = bd->free;
+ stp->n_blocks -= 1;
+ stp->n_words -= bd->free - bd->start;
+ t++;
+ if (t == n_gc_threads) t = 0;
+ }
+ }
+
+
+ // Move the private mutable lists from each capability onto the
+ // main mutable list for the generation.
+ for (i = 0; i < n_capabilities; i++) {
+ for (bd = capabilities[i].mut_lists[g];
+ bd->link != NULL; bd = bd->link) {
+ /* nothing */
+ }
+ bd->link = generations[g].mut_list;
+ generations[g].mut_list = capabilities[i].mut_lists[g];
+ capabilities[i].mut_lists[g] = allocBlock();
+ }
+}
+
+/* -----------------------------------------------------------------------------
+ Initialise a gc_thread before GC
+ -------------------------------------------------------------------------- */
+
+static void
+init_gc_thread (gc_thread *t)
+{
+ t->static_objects = END_OF_STATIC_LIST;
+ t->scavenged_static_objects = END_OF_STATIC_LIST;
+ t->scan_bd = NULL;
+ t->evac_step = 0;
+ t->failed_to_evac = rtsFalse;
+ t->eager_promotion = rtsTrue;
+ t->thunk_selector_depth = 0;
+ t->copied = 0;
+ t->scanned = 0;
+ t->any_work = 0;
+ t->no_work = 0;
+ t->scav_find_work = 0;