+/* -----------------------------------------------------------------------------
+ Figure out which generation to collect, initialise N and major_gc.
+ -------------------------------------------------------------------------- */
+
+static void
+initialise_N (rtsBool force_major_gc)
+{
+ nat g;
+
+ if (force_major_gc) {
+ N = RtsFlags.GcFlags.generations - 1;
+ major_gc = rtsTrue;
+ } else {
+ N = 0;
+ for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
+ if (generations[g].steps[0].n_blocks +
+ generations[g].steps[0].n_large_blocks
+ >= generations[g].max_blocks) {
+ N = g;
+ }
+ }
+ major_gc = (N == RtsFlags.GcFlags.generations-1);
+ }
+}
+
+/* -----------------------------------------------------------------------------
+ Initialise the gc_thread structures.
+ -------------------------------------------------------------------------- */
+
+static void
+alloc_gc_thread (gc_thread *t, int n)
+{
+ nat g, s;
+ step_workspace *ws;
+
+ t->thread_index = n;
+ t->free_blocks = NULL;
+ t->gc_count = 0;
+
+ init_gc_thread(t);
+
+ t->steps = stgMallocBytes(RtsFlags.GcFlags.generations *
+ sizeof(step_workspace *),
+ "initialise_gc_thread");
+
+ for (g = 0; g < RtsFlags.GcFlags.generations; g++)
+ {
+ t->steps[g] = stgMallocBytes(generations[g].n_steps *
+ sizeof(step_workspace),
+ "initialise_gc_thread/2");
+
+ for (s = 0; s < generations[g].n_steps; s++)
+ {
+ ws = &t->steps[g][s];
+ ws->stp = &generations[g].steps[s];
+ ws->gct = t;
+
+ ws->scan_bd = NULL;
+ ws->scan = NULL;
+
+ ws->todo_bd = NULL;
+ ws->buffer_todo_bd = NULL;
+
+ ws->scavd_list = NULL;
+ ws->n_scavd_blocks = 0;
+ }
+ }
+}
+
+
+static void
+alloc_gc_threads (void)
+{
+ if (gc_threads == NULL) {
+#if defined(THREADED_RTS)
+ nat i;
+
+ gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
+ sizeof(gc_thread),
+ "alloc_gc_threads");
+
+ for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
+ alloc_gc_thread(&gc_threads[i], i);
+ }
+#else
+ gc_threads = stgMallocBytes (sizeof(gc_thread),
+ "alloc_gc_threads");
+
+ alloc_gc_thread(gc_threads, 0);
+#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++) {
+
+ // generation 0, step 0 doesn't need to-space
+ if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
+ continue;
+ }
+
+ stp = &generations[g].steps[s];
+ ASSERT(stp->gen_no == g);
+
+ // deprecate the existing blocks
+ stp->old_blocks = stp->blocks;
+ stp->n_old_blocks = stp->n_blocks;
+ stp->blocks = NULL;
+ stp->n_blocks = 0;
+
+ // we don't have any to-be-scavenged blocks yet
+ stp->todos = NULL;
+ stp->n_todos = 0;
+
+ // initialise the large object queues.
+ stp->scavenged_large_objects = NULL;
+ stp->n_scavenged_large_blocks = 0;
+
+ // mark the large objects as not evacuated yet
+ 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->is_compacted) {
+ 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_COMPACTED flag
+ // for this block. The invariant is that
+ // BF_COMPACTED is always unset, except during GC
+ // when it is set on those blocks which will be
+ // compacted.
+ bd->flags |= BF_COMPACTED;
+ }
+ }
+ }
+ }
+
+ // 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][s];
+
+ ws->scan_bd = NULL;
+ ws->scan = NULL;
+
+ ws->todo_large_objects = NULL;
+
+ // allocate the first to-space block; extra blocks will be
+ // chained on as necessary.
+ ws->todo_bd = NULL;
+ ws->buffer_todo_bd = NULL;
+ gc_alloc_todo_block(ws);
+
+ // allocate a block for "already scavenged" objects. This goes
+ // on the front of the stp->blocks list, so it won't be
+ // traversed by the scavenging sweep.
+ ws->scavd_list = NULL;
+ ws->n_scavd_blocks = 0;
+ gc_alloc_scavd_block(ws);
+ }
+ }
+}
+
+
+/* ----------------------------------------------------------------------------
+ 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 (t = 0; t < threads; t++) {
+ for (s = 0; s < generations[g].n_steps; s++) {
+
+ ws = &gc_threads[t].steps[g][s];
+ stp = ws->stp;
+
+ ws->buffer_todo_bd = NULL;
+ ws->todo_large_objects = NULL;
+
+ // 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 (isPartiallyFull(stp->blocks))
+ {
+ ws->todo_bd = stp->blocks;
+ stp->blocks = stp->blocks->link;
+ stp->n_blocks -= 1;
+ ws->todo_bd->link = NULL;
+
+ // this block is also the scan block; we must scan
+ // from the current end point.
+ ws->scan_bd = ws->todo_bd;
+ ws->scan = ws->scan_bd->free;
+
+ // subtract the contents of this block from the stats,
+ // because we'll count the whole block later.
+ copied -= ws->scan_bd->free - ws->scan_bd->start;
+ }
+ else
+ {
+ ws->scan_bd = NULL;
+ ws->scan = NULL;
+ ws->todo_bd = NULL;
+ gc_alloc_todo_block(ws);
+ }
+
+ // Do the same trick for the scavd block
+ if (isPartiallyFull(stp->blocks))
+ {
+ ws->scavd_list = stp->blocks;
+ stp->blocks = stp->blocks->link;
+ stp->n_blocks -= 1;
+ ws->scavd_list->link = NULL;
+ ws->n_scavd_blocks = 1;
+ // subtract the contents of this block from the stats,
+ // because we'll count the whole block later.
+ scavd_copied -= ws->scavd_list->free - ws->scavd_list->start;
+ }
+ else
+ {
+ ws->scavd_list = NULL;
+ ws->n_scavd_blocks = 0;
+ gc_alloc_scavd_block(ws);
+ }
+ }
+ }
+
+ // 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->evac_gen = 0;
+ t->failed_to_evac = rtsFalse;
+ t->eager_promotion = rtsTrue;
+ t->thunk_selector_depth = 0;
+}
+
+/* -----------------------------------------------------------------------------
+ Function we pass to GetRoots to evacuate roots.
+ -------------------------------------------------------------------------- */
+