/* Thread-local data for each GC thread
*/
-gc_thread *gc_threads = NULL;
+gc_thread **gc_threads = NULL;
// gc_thread *gct = NULL; // this thread's gct TODO: make thread-local
// Number of threads running in *this* GC. Affects how many
}
#endif
- // tell the STM to discard any cached closures it's hoping to re-use
- stmPreGCHook();
-
// tell the stats department that we've started a GC
stat_startGC();
-#ifdef DEBUG
- // check for memory leaks if DEBUG is on
- memInventory();
-#endif
+ // tell the STM to discard any cached closures it's hoping to re-use
+ stmPreGCHook();
#ifdef DEBUG
mutlist_MUTVARS = 0;
}
#endif
+#ifdef DEBUG
+ // check for memory leaks if DEBUG is on
+ memInventory(traceClass(DEBUG_gc));
+#endif
+
// check stack sanity *before* GC (ToDo: check all threads)
IF_DEBUG(sanity, checkFreeListSanity());
static_objects = END_OF_STATIC_LIST;
scavenged_static_objects = END_OF_STATIC_LIST;
-#ifdef THREADED_RTS
- initSpinLock(&static_objects_sync);
- initSpinLock(&recordMutableGen_sync);
- initSpinLock(&gc_alloc_block_sync);
-#endif
-
// Initialise all the generations/steps that we're collecting.
for (g = 0; g <= N; g++) {
init_collected_gen(g,n_gc_threads);
// Initialise all our gc_thread structures
for (t = 0; t < n_gc_threads; t++) {
- init_gc_thread(&gc_threads[t]);
+ init_gc_thread(gc_threads[t]);
}
// the main thread is running: this prevents any other threads from
copied = 0;
// this is the main thread
- gct = &gc_threads[0];
+ gct = gc_threads[0];
/* -----------------------------------------------------------------------
* follow all the roots that we know about:
bdescr *prev;
for (t = 0; t < n_gc_threads; t++) {
- thr = &gc_threads[t];
-
- for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
- for (s = 0; s < generations[g].n_steps; s++) {
- ws = &thr->steps[g][s];
- if (g==0 && s==0) continue;
-
- // Not true?
- // ASSERT( ws->scan_bd == ws->todo_bd );
- ASSERT( ws->scan_bd ? ws->scan == ws->scan_bd->free : 1 );
-
- // Push the final block
- if (ws->scan_bd) { push_scan_block(ws->scan_bd, ws); }
-
- ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
-
- prev = ws->scavd_list;
- for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
- bd->flags &= ~BF_EVACUATED; // now from-space
- prev = bd;
- }
- prev->link = ws->stp->blocks;
- ws->stp->blocks = ws->scavd_list;
- ws->stp->n_blocks += ws->n_scavd_blocks;
- ASSERT(countBlocks(ws->stp->blocks) == ws->stp->n_blocks);
- }
+ thr = gc_threads[t];
+
+ // not step 0
+ for (s = 1; s < total_steps; s++) {
+ ws = &thr->steps[s];
+ // Not true?
+ // ASSERT( ws->scan_bd == ws->todo_bd );
+ ASSERT( ws->scan_bd ? ws->scan == ws->scan_bd->free : 1 );
+
+ // Push the final block
+ if (ws->scan_bd) { push_scan_block(ws->scan_bd, ws); }
+
+ ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
+
+ prev = ws->scavd_list;
+ for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
+ bd->flags &= ~BF_EVACUATED; // now from-space
+ prev = bd;
+ }
+ prev->link = ws->stp->blocks;
+ ws->stp->blocks = ws->scavd_list;
+ ws->stp->n_blocks += ws->n_scavd_blocks;
+ ASSERT(countBlocks(ws->stp->blocks) == ws->stp->n_blocks);
}
}
}
resize_generations();
// Guess the amount of live data for stats.
- live = calcLive();
+ live = calcLiveBlocks() * BLOCK_SIZE_W;
+ debugTrace(DEBUG_gc, "Slop: %ldKB",
+ (live - calcLiveWords()) / (1024/sizeof(W_)));
// Free the small objects allocated via allocate(), since this will
// all have been copied into G0S1 now.
#ifdef DEBUG
// check for memory leaks if DEBUG is on
- memInventory();
+ memInventory(traceClass(DEBUG_gc));
#endif
#ifdef RTS_GTK_FRONTPANEL
gct = saved_gct;
}
+/* -----------------------------------------------------------------------------
+ * Mark all nodes pointed to by sparks in the spark queues (for GC) Does an
+ * implicit slide i.e. after marking all sparks are at the beginning of the
+ * spark pool and the spark pool only contains sparkable closures
+ * -------------------------------------------------------------------------- */
+
+#ifdef THREADED_RTS
+static void
+markSparkQueue (evac_fn evac, Capability *cap)
+{
+ StgClosure **sparkp, **to_sparkp;
+ nat n, pruned_sparks; // stats only
+ StgSparkPool *pool;
+
+ PAR_TICKY_MARK_SPARK_QUEUE_START();
+
+ n = 0;
+ pruned_sparks = 0;
+
+ pool = &(cap->r.rSparks);
+
+ ASSERT_SPARK_POOL_INVARIANTS(pool);
+
+#if defined(PARALLEL_HASKELL)
+ // stats only
+ n = 0;
+ pruned_sparks = 0;
+#endif
+
+ sparkp = pool->hd;
+ to_sparkp = pool->hd;
+ while (sparkp != pool->tl) {
+ ASSERT(*sparkp!=NULL);
+ ASSERT(LOOKS_LIKE_CLOSURE_PTR(((StgClosure *)*sparkp)));
+ // ToDo?: statistics gathering here (also for GUM!)
+ if (closure_SHOULD_SPARK(*sparkp)) {
+ evac(sparkp);
+ *to_sparkp++ = *sparkp;
+ if (to_sparkp == pool->lim) {
+ to_sparkp = pool->base;
+ }
+ n++;
+ } else {
+ pruned_sparks++;
+ }
+ sparkp++;
+ if (sparkp == pool->lim) {
+ sparkp = pool->base;
+ }
+ }
+ pool->tl = to_sparkp;
+
+ PAR_TICKY_MARK_SPARK_QUEUE_END(n);
+
+#if defined(PARALLEL_HASKELL)
+ debugTrace(DEBUG_sched,
+ "marked %d sparks and pruned %d sparks on [%x]",
+ n, pruned_sparks, mytid);
+#else
+ debugTrace(DEBUG_sched,
+ "marked %d sparks and pruned %d sparks",
+ n, pruned_sparks);
+#endif
+
+ debugTrace(DEBUG_sched,
+ "new spark queue len=%d; (hd=%p; tl=%p)\n",
+ sparkPoolSize(pool), pool->hd, pool->tl);
+}
+#endif
+
/* ---------------------------------------------------------------------------
Where are the roots that we know about?
Initialise the gc_thread structures.
-------------------------------------------------------------------------- */
-static void
-alloc_gc_thread (gc_thread *t, int n)
+static gc_thread *
+alloc_gc_thread (int n)
{
- nat g, s;
+ nat s;
step_workspace *ws;
+ gc_thread *t;
+
+ t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
+ "alloc_gc_thread");
#ifdef THREADED_RTS
t->id = 0;
t->papi_events = -1;
#endif
- t->steps = stgMallocBytes(RtsFlags.GcFlags.generations *
- sizeof(step_workspace *),
- "initialise_gc_thread");
-
- for (g = 0; g < RtsFlags.GcFlags.generations; g++)
+ for (s = 0; s < total_steps; s++)
{
- 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;
- }
+ ws = &t->steps[s];
+ ws->stp = &all_steps[s];
+ ASSERT(s == ws->stp->abs_no);
+ 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;
}
+
+ return t;
}
#if defined(THREADED_RTS)
nat i;
gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
- sizeof(gc_thread),
+ sizeof(gc_thread*),
"alloc_gc_threads");
for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
- alloc_gc_thread(&gc_threads[i], i);
+ gc_threads[i] = alloc_gc_thread(i);
}
#else
- gc_threads = stgMallocBytes (sizeof(gc_thread),
+ gc_threads = stgMallocBytes (sizeof(gc_thread*),
"alloc_gc_threads");
- alloc_gc_thread(gc_threads, 0);
+ gc_threads[0] = alloc_gc_thread(0);
#endif
}
}
// 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]);
+ gc_threads[i]);
}
done = rtsTrue;
}
nat i;
for (i=1; i < n_threads; i++) {
inc_running();
- ACQUIRE_LOCK(&gc_threads[i].wake_mutex);
- gc_threads[i].wakeup = rtsTrue;
- signalCondition(&gc_threads[i].wake_cond);
- RELEASE_LOCK(&gc_threads[i].wake_mutex);
+ ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
+ gc_threads[i]->wakeup = rtsTrue;
+ signalCondition(&gc_threads[i]->wake_cond);
+ RELEASE_LOCK(&gc_threads[i]->wake_mutex);
}
#endif
}
// 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 = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
ws->scan_bd = NULL;
ws->scan = NULL;
for (t = 0; t < threads; t++) {
for (s = 0; s < generations[g].n_steps; s++) {
- ws = &gc_threads[t].steps[g][s];
+ ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
stp = ws->stp;
ws->buffer_todo_bd = 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))
+ if (stp->blocks && isPartiallyFull(stp->blocks))
{
ws->todo_bd = stp->blocks;
ws->todo_free = ws->todo_bd->free;
projects / ghc-hetmet.git / blobdiff