*/
static nat evac_gen;
+/* Whether to do eager promotion or not.
+ */
+static rtsBool eager_promotion;
+
/* Weak pointers
*/
StgWeak *old_weak_ptr_list; // also pending finaliser list
static lnat thunk_selector_depth = 0;
#define MAX_THUNK_SELECTOR_DEPTH 8
+/* Mut-list stats */
+#ifdef DEBUG
+static nat
+ mutlist_MUTVARS,
+ mutlist_MUTARRS,
+ mutlist_OTHERS;
+#endif
+
/* -----------------------------------------------------------------------------
Static function declarations
-------------------------------------------------------------------------- */
{
bdescr *bd;
step *stp;
- lnat live, allocated, collected = 0, copied = 0, scavd_copied = 0;
+ lnat live, allocated, copied = 0, scavd_copied = 0;
lnat oldgen_saved_blocks = 0;
nat g, s, i;
memInventory();
#endif
+#ifdef DEBUG
+ mutlist_MUTVARS = 0;
+ mutlist_MUTARRS = 0;
+ mutlist_OTHERS = 0;
+#endif
+
// Init stats and print par specific (timing) info
PAR_TICKY_PAR_START();
mark_stack_bdescr = NULL;
}
+ eager_promotion = rtsTrue; // for now
+
/* -----------------------------------------------------------------------
* follow all the roots that we know about:
* - mutable lists from each generation > N
// Count the mutable list as bytes "copied" for the purposes of
// stats. Every mutable list is copied during every GC.
if (g > 0) {
+ nat mut_list_size = 0;
for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
- copied += (bd->free - bd->start) * sizeof(StgWord);
+ mut_list_size += bd->free - bd->start;
}
+ copied += mut_list_size;
+
+ IF_DEBUG(gc, debugBelch("mut_list_size: %ld (%d vars, %d arrays, %d others)\n", mut_list_size * sizeof(W_), mutlist_MUTVARS, mutlist_MUTARRS, mutlist_OTHERS));
}
for (s = 0; s < generations[g].n_steps; s++) {
// for generations we collected...
if (g <= N) {
- // rough calculation of garbage collected, for stats output
- if (stp->is_compacted) {
- collected += (oldgen_saved_blocks - stp->n_old_blocks) * BLOCK_SIZE_W;
- } else {
- if (g == 0 && s == 0) {
- collected += countNurseryBlocks() * BLOCK_SIZE_W;
- collected += alloc_blocks;
- } else {
- collected += stp->n_old_blocks * BLOCK_SIZE_W;
- }
- }
-
/* free old memory and shift to-space into from-space for all
* the collected steps (except the allocation area). These
* freed blocks will probaby be quickly recycled.
ACQUIRE_SM_LOCK;
// send exceptions to any threads which were about to die
+ RELEASE_SM_LOCK;
resurrectThreads(resurrected_threads);
+ ACQUIRE_SM_LOCK;
// Update the stable pointer hash table.
updateStablePtrTable(major_gc);
#endif
// ok, GC over: tell the stats department what happened.
- stat_endGC(allocated, collected, live, copied, scavd_copied, N);
+ stat_endGC(allocated, live, copied, scavd_copied, N);
#if defined(RTS_USER_SIGNALS)
// unblock signals again
* by evacuate()).
*/
if (stp->gen_no < evac_gen) {
-#ifdef NO_EAGER_PROMOTION
- failed_to_evac = rtsTrue;
-#else
- stp = &generations[evac_gen].steps[0];
-#endif
+ if (eager_promotion) {
+ stp = &generations[evac_gen].steps[0];
+ } else {
+ failed_to_evac = rtsTrue;
+ }
}
/* chain a new block onto the to-space for the destination step if
* by evacuate()).
*/
if (stp->gen_no < evac_gen) {
-#ifdef NO_EAGER_PROMOTION
- failed_to_evac = rtsTrue;
-#else
- stp = &generations[evac_gen].steps[0];
-#endif
+ if (eager_promotion) {
+ stp = &generations[evac_gen].steps[0];
+ } else {
+ failed_to_evac = rtsTrue;
+ }
}
/* chain a new block onto the to-space for the destination step if
TICK_GC_WORDS_COPIED(size_to_copy);
if (stp->gen_no < evac_gen) {
-#ifdef NO_EAGER_PROMOTION
- failed_to_evac = rtsTrue;
-#else
- stp = &generations[evac_gen].steps[0];
-#endif
+ if (eager_promotion) {
+ stp = &generations[evac_gen].steps[0];
+ } else {
+ failed_to_evac = rtsTrue;
+ }
}
if (stp->hp + size_to_reserve >= stp->hpLim) {
SET_EVACUAEE_FOR_LDV(src, size_to_reserve);
// fill the slop
if (size_to_reserve - size_to_copy_org > 0)
- FILL_SLOP(stp->hp - 1, (int)(size_to_reserve - size_to_copy_org));
+ LDV_FILL_SLOP(stp->hp - 1, (int)(size_to_reserve - size_to_copy_org));
#endif
return (StgClosure *)dest;
}
*/
stp = bd->step->to;
if (stp->gen_no < evac_gen) {
-#ifdef NO_EAGER_PROMOTION
- failed_to_evac = rtsTrue;
-#else
- stp = &generations[evac_gen].steps[0];
-#endif
+ if (eager_promotion) {
+ stp = &generations[evac_gen].steps[0];
+ } else {
+ failed_to_evac = rtsTrue;
+ }
}
bd->step = stp;
switch (info->type) {
- case MUT_VAR:
+ case MUT_VAR_CLEAN:
+ case MUT_VAR_DIRTY:
case MVAR:
return copy(q,sizeW_fromITBL(info),stp);
case THUNK_SELECTOR:
{
StgClosure *p;
+ const StgInfoTable *info_ptr;
if (thunk_selector_depth > MAX_THUNK_SELECTOR_DEPTH) {
return copy(q,THUNK_SELECTOR_sizeW(),stp);
}
+ // stashed away for LDV profiling, see below
+ info_ptr = q->header.info;
+
p = eval_thunk_selector(info->layout.selector_offset,
(StgSelector *)q);
val = evacuate(p);
thunk_selector_depth--;
+#ifdef PROFILING
+ // For the purposes of LDV profiling, we have destroyed
+ // the original selector thunk.
+ SET_INFO(q, info_ptr);
+ LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(q);
+#endif
+
// Update the THUNK_SELECTOR with an indirection to the
// EVACUATED closure now at p. Why do this rather than
// upd_evacuee(q,p)? Because we have an invariant that an
SET_INFO(q, &stg_IND_info);
((StgInd *)q)->indirectee = p;
-#ifdef PROFILING
- // We store the size of the just evacuated object in the
- // LDV word so that the profiler can guess the position of
- // the next object later.
- SET_EVACUAEE_FOR_LDV(q, THUNK_SELECTOR_sizeW());
-#endif
+ // For the purposes of LDV profiling, we have created an
+ // indirection.
+ LDV_RECORD_CREATE(q);
+
return val;
}
}
// just copy the block
return copy_noscav(q,arr_words_sizeW((StgArrWords *)q),stp);
- case MUT_ARR_PTRS:
+ case MUT_ARR_PTRS_CLEAN:
+ case MUT_ARR_PTRS_DIRTY:
case MUT_ARR_PTRS_FROZEN:
case MUT_ARR_PTRS_FROZEN0:
// just copy the block
}
case BLOCKED_FETCH:
- ASSERT(sizeofW(StgBlockedFetch) >= MIN_NONUPD_SIZE);
+ ASSERT(sizeofW(StgBlockedFetch) >= MIN_PAYLOD_SIZE);
to = copy(q,sizeofW(StgBlockedFetch),stp);
IF_DEBUG(gc,
debugBelch("@@ evacuate: %p (%s) to %p (%s)",
case REMOTE_REF:
# endif
case FETCH_ME:
- ASSERT(sizeofW(StgBlockedFetch) >= MIN_UPD_SIZE);
+ ASSERT(sizeofW(StgBlockedFetch) >= MIN_PAYLOAD_SIZE);
to = copy(q,sizeofW(StgFetchMe),stp);
IF_DEBUG(gc,
debugBelch("@@ evacuate: %p (%s) to %p (%s)",
return to;
case FETCH_ME_BQ:
- ASSERT(sizeofW(StgBlockedFetch) >= MIN_UPD_SIZE);
+ ASSERT(sizeofW(StgBlockedFetch) >= MIN_PAYLOAD_SIZE);
to = copy(q,sizeofW(StgFetchMeBlockingQueue),stp);
IF_DEBUG(gc,
debugBelch("@@ evacuate: %p (%s) to %p (%s)",
p += sizeofW(StgInd);
break;
- case MUT_VAR:
- evac_gen = 0;
+ case MUT_VAR_CLEAN:
+ case MUT_VAR_DIRTY: {
+ rtsBool saved_eager_promotion = eager_promotion;
+
+ eager_promotion = rtsFalse;
((StgMutVar *)p)->var = evacuate(((StgMutVar *)p)->var);
- evac_gen = saved_evac_gen;
- failed_to_evac = rtsTrue; // mutable anyhow
+ eager_promotion = saved_eager_promotion;
+
+ if (failed_to_evac) {
+ ((StgClosure *)q)->header.info = &stg_MUT_VAR_DIRTY_info;
+ } else {
+ ((StgClosure *)q)->header.info = &stg_MUT_VAR_CLEAN_info;
+ }
p += sizeofW(StgMutVar);
break;
+ }
case CAF_BLACKHOLE:
case SE_CAF_BLACKHOLE:
p += arr_words_sizeW((StgArrWords *)p);
break;
- case MUT_ARR_PTRS:
+ case MUT_ARR_PTRS_CLEAN:
+ case MUT_ARR_PTRS_DIRTY:
// follow everything
{
StgPtr next;
-
- evac_gen = 0; // repeatedly mutable
+ rtsBool saved_eager;
+
+ // We don't eagerly promote objects pointed to by a mutable
+ // array, but if we find the array only points to objects in
+ // the same or an older generation, we mark it "clean" and
+ // avoid traversing it during minor GCs.
+ saved_eager = eager_promotion;
+ eager_promotion = rtsFalse;
next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
*p = (StgWord)(StgPtr)evacuate((StgClosure *)*p);
}
- evac_gen = saved_evac_gen;
- failed_to_evac = rtsTrue; // mutable anyhow.
+ eager_promotion = saved_eager;
+
+ if (failed_to_evac) {
+ ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_DIRTY_info;
+ } else {
+ ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_CLEAN_info;
+ }
+
+ failed_to_evac = rtsTrue; // always put it on the mutable list.
break;
}
for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
*p = (StgWord)(StgPtr)evacuate((StgClosure *)*p);
}
- // it's tempting to recordMutable() if failed_to_evac is
- // false, but that breaks some assumptions (eg. every
- // closure on the mutable list is supposed to have the MUT
- // flag set, and MUT_ARR_PTRS_FROZEN doesn't).
+
+ // If we're going to put this object on the mutable list, then
+ // set its info ptr to MUT_ARR_PTRS_FROZEN0 to indicate that.
+ if (failed_to_evac) {
+ ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN0_info;
+ } else {
+ ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN_info;
+ }
break;
}
case TSO:
{
StgTSO *tso = (StgTSO *)p;
- evac_gen = 0;
+ rtsBool saved_eager = eager_promotion;
+
+ eager_promotion = rtsFalse;
scavengeTSO(tso);
- evac_gen = saved_evac_gen;
- failed_to_evac = rtsTrue; // mutable anyhow.
+ eager_promotion = saved_eager;
+
+ if (failed_to_evac) {
+ tso->flags |= TSO_DIRTY;
+ } else {
+ tso->flags &= ~TSO_DIRTY;
+ }
+
+ failed_to_evac = rtsTrue; // always on the mutable list
p += tso_sizeW(tso);
break;
}
evac_gen = 0;
tvar->current_value = evacuate((StgClosure*)tvar->current_value);
tvar->first_wait_queue_entry = (StgTVarWaitQueue *)evacuate((StgClosure*)tvar->first_wait_queue_entry);
-#if defined(SMP)
- tvar->last_update_by = (StgTRecHeader *)evacuate((StgClosure*)tvar->last_update_by);
-#endif
evac_gen = saved_evac_gen;
failed_to_evac = rtsTrue; // mutable
p += sizeofW(StgTVar);
evacuate(((StgInd *)p)->indirectee);
break;
- case MUT_VAR:
- evac_gen = 0;
+ case MUT_VAR_CLEAN:
+ case MUT_VAR_DIRTY: {
+ rtsBool saved_eager_promotion = eager_promotion;
+
+ eager_promotion = rtsFalse;
((StgMutVar *)p)->var = evacuate(((StgMutVar *)p)->var);
- evac_gen = saved_evac_gen;
- failed_to_evac = rtsTrue;
+ eager_promotion = saved_eager_promotion;
+
+ if (failed_to_evac) {
+ ((StgClosure *)q)->header.info = &stg_MUT_VAR_DIRTY_info;
+ } else {
+ ((StgClosure *)q)->header.info = &stg_MUT_VAR_CLEAN_info;
+ }
break;
+ }
case CAF_BLACKHOLE:
case SE_CAF_BLACKHOLE:
scavenge_AP((StgAP *)p);
break;
- case MUT_ARR_PTRS:
+ case MUT_ARR_PTRS_CLEAN:
+ case MUT_ARR_PTRS_DIRTY:
// follow everything
{
StgPtr next;
-
- evac_gen = 0; // repeatedly mutable
+ rtsBool saved_eager;
+
+ // We don't eagerly promote objects pointed to by a mutable
+ // array, but if we find the array only points to objects in
+ // the same or an older generation, we mark it "clean" and
+ // avoid traversing it during minor GCs.
+ saved_eager = eager_promotion;
+ eager_promotion = rtsFalse;
next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
*p = (StgWord)(StgPtr)evacuate((StgClosure *)*p);
}
- evac_gen = saved_evac_gen;
+ eager_promotion = saved_eager;
+
+ if (failed_to_evac) {
+ ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_DIRTY_info;
+ } else {
+ ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_CLEAN_info;
+ }
+
failed_to_evac = rtsTrue; // mutable anyhow.
break;
}
case MUT_ARR_PTRS_FROZEN0:
// follow everything
{
- StgPtr next;
+ StgPtr next, q = p;
next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
*p = (StgWord)(StgPtr)evacuate((StgClosure *)*p);
}
+
+ // If we're going to put this object on the mutable list, then
+ // set its info ptr to MUT_ARR_PTRS_FROZEN0 to indicate that.
+ if (failed_to_evac) {
+ ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN0_info;
+ } else {
+ ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN_info;
+ }
break;
}
case TSO:
{
StgTSO *tso = (StgTSO *)p;
- evac_gen = 0;
+ rtsBool saved_eager = eager_promotion;
+
+ eager_promotion = rtsFalse;
scavengeTSO(tso);
- evac_gen = saved_evac_gen;
- failed_to_evac = rtsTrue;
+ eager_promotion = saved_eager;
+
+ if (failed_to_evac) {
+ tso->flags |= TSO_DIRTY;
+ } else {
+ tso->flags &= ~TSO_DIRTY;
+ }
+
+ failed_to_evac = rtsTrue; // always on the mutable list
break;
}
evac_gen = 0;
tvar->current_value = evacuate((StgClosure*)tvar->current_value);
tvar->first_wait_queue_entry = (StgTVarWaitQueue *)evacuate((StgClosure*)tvar->first_wait_queue_entry);
-#if defined(SMP)
- tvar->last_update_by = (StgTRecHeader *)evacuate((StgClosure*)tvar->last_update_by);
-#endif
evac_gen = saved_evac_gen;
failed_to_evac = rtsTrue; // mutable
break;
// already scavenged?
if (is_marked(oldgen_scan+1,oldgen_scan_bd)) {
- oldgen_scan += sizeofW(StgHeader) + MIN_NONUPD_SIZE;
+ oldgen_scan += sizeofW(StgHeader) + MIN_PAYLOAD_SIZE;
goto loop;
}
push_mark_stack(oldgen_scan);
// ToDo: bump the linear scan by the actual size of the object
- oldgen_scan += sizeofW(StgHeader) + MIN_NONUPD_SIZE;
+ oldgen_scan += sizeofW(StgHeader) + MIN_PAYLOAD_SIZE;
goto linear_scan;
}
break;
}
- case MUT_VAR:
- evac_gen = 0;
+ case MUT_VAR_CLEAN:
+ case MUT_VAR_DIRTY: {
+ StgPtr q = p;
+ rtsBool saved_eager_promotion = eager_promotion;
+
+ eager_promotion = rtsFalse;
((StgMutVar *)p)->var = evacuate(((StgMutVar *)p)->var);
- evac_gen = saved_evac_gen;
- failed_to_evac = rtsTrue; // mutable anyhow
+ eager_promotion = saved_eager_promotion;
+
+ if (failed_to_evac) {
+ ((StgClosure *)q)->header.info = &stg_MUT_VAR_DIRTY_info;
+ } else {
+ ((StgClosure *)q)->header.info = &stg_MUT_VAR_CLEAN_info;
+ }
break;
+ }
case CAF_BLACKHOLE:
case SE_CAF_BLACKHOLE:
// nothing to follow
break;
- case MUT_ARR_PTRS:
+ case MUT_ARR_PTRS_CLEAN:
+ case MUT_ARR_PTRS_DIRTY:
{
- // follow everything
- StgPtr next;
-
- evac_gen = 0; // repeatedly mutable
+ StgPtr next, q;
+ rtsBool saved_eager;
+
+ // We don't eagerly promote objects pointed to by a mutable
+ // array, but if we find the array only points to objects in
+ // the same or an older generation, we mark it "clean" and
+ // avoid traversing it during minor GCs.
+ saved_eager = eager_promotion;
+ eager_promotion = rtsFalse;
+ q = p;
next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
*p = (StgWord)(StgPtr)evacuate((StgClosure *)*p);
}
- evac_gen = saved_evac_gen;
+ eager_promotion = saved_eager;
+
+ if (failed_to_evac) {
+ ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_DIRTY_info;
+ } else {
+ ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_CLEAN_info;
+ }
+
failed_to_evac = rtsTrue;
break;
}
case MUT_ARR_PTRS_FROZEN0:
{
// follow everything
- StgPtr next;
+ StgPtr next, q=p;
next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
*p = (StgWord)(StgPtr)evacuate((StgClosure *)*p);
}
+
+ // If we're going to put this object on the mutable list, then
+ // set its info ptr to MUT_ARR_PTRS_FROZEN0 to indicate that.
+ if (failed_to_evac) {
+ ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN0_info;
+ } else {
+ ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN_info;
+ }
break;
}
case TSO:
{
StgTSO *tso = (StgTSO *)p;
-
- evac_gen = 0; // repeatedly mutable
+ rtsBool saved_eager = eager_promotion;
+
+ eager_promotion = rtsFalse;
scavengeTSO(tso);
- evac_gen = saved_evac_gen;
- failed_to_evac = rtsTrue;
+ eager_promotion = saved_eager;
+
+ if (failed_to_evac) {
+ tso->flags |= TSO_DIRTY;
+ } else {
+ tso->flags &= ~TSO_DIRTY;
+ }
+
+ failed_to_evac = rtsTrue; // always on the mutable list
break;
}
evac_gen = 0;
tvar->current_value = evacuate((StgClosure*)tvar->current_value);
tvar->first_wait_queue_entry = (StgTVarWaitQueue *)evacuate((StgClosure*)tvar->first_wait_queue_entry);
-#if defined(SMP)
- tvar->last_update_by = (StgTRecHeader *)evacuate((StgClosure*)tvar->last_update_by);
-#endif
evac_gen = saved_evac_gen;
failed_to_evac = rtsTrue; // mutable
break;
for (q = bd->start; q < bd->free; q++) {
p = (StgPtr)*q;
ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
+
+#ifdef DEBUG
+ switch (get_itbl((StgClosure *)p)->type) {
+ case MUT_VAR_CLEAN:
+ barf("MUT_VAR_CLEAN on mutable list");
+ case MUT_VAR_DIRTY:
+ mutlist_MUTVARS++; break;
+ case MUT_ARR_PTRS_CLEAN:
+ case MUT_ARR_PTRS_DIRTY:
+ case MUT_ARR_PTRS_FROZEN:
+ case MUT_ARR_PTRS_FROZEN0:
+ mutlist_MUTARRS++; break;
+ default:
+ mutlist_OTHERS++; break;
+ }
+#endif
+
+ // Check whether this object is "clean", that is it
+ // definitely doesn't point into a young generation.
+ // Clean objects don't need to be scavenged. Some clean
+ // objects (MUT_VAR_CLEAN) are not kept on the mutable
+ // list at all; others, such as MUT_ARR_PTRS_CLEAN and
+ // TSO, are always on the mutable list.
+ //
+ switch (get_itbl((StgClosure *)p)->type) {
+ case MUT_ARR_PTRS_CLEAN:
+ recordMutableGen((StgClosure *)p,gen);
+ continue;
+ case TSO: {
+ StgTSO *tso = (StgTSO *)p;
+ if ((tso->flags & TSO_DIRTY) == 0) {
+ // A clean TSO: we don't have to traverse its
+ // stack. However, we *do* follow the link field:
+ // we don't want to have to mark a TSO dirty just
+ // because we put it on a different queue.
+ if (tso->why_blocked != BlockedOnBlackHole) {
+ tso->link = (StgTSO *)evacuate((StgClosure *)tso->link);
+ }
+ recordMutableGen((StgClosure *)p,gen);
+ continue;
+ }
+ }
+ default:
+ ;
+ }
+
if (scavenge_one(p)) {
- /* didn't manage to promote everything, so put the
- * object back on the list.
- */
+ // didn't manage to promote everything, so put the
+ // object back on the list.
recordMutableGen((StgClosure *)p,gen);
}
}
switch (info->i.type) {
case UPDATE_FRAME:
+ // In SMP, we can get update frames that point to indirections
+ // when two threads evaluate the same thunk. We do attempt to
+ // discover this situation in threadPaused(), but it's
+ // possible that the following sequence occurs:
+ //
+ // A B
+ // enter T
+ // enter T
+ // blackhole T
+ // update T
+ // GC
+ //
+ // Now T is an indirection, and the update frame is already
+ // marked on A's stack, so we won't traverse it again in
+ // threadPaused(). We could traverse the whole stack again
+ // before GC, but that seems like overkill.
+ //
+ // Scavenging this update frame as normal would be disastrous;
+ // the updatee would end up pointing to the value. So we turn
+ // the indirection into an IND_PERM, so that evacuate will
+ // copy the indirection into the old generation instead of
+ // discarding it.
+ if (get_itbl(((StgUpdateFrame *)p)->updatee)->type == IND) {
+ ((StgUpdateFrame *)p)->updatee->header.info =
+ (StgInfoTable *)&stg_IND_PERM_info;
+ }
((StgUpdateFrame *)p)->updatee
= evacuate(((StgUpdateFrame *)p)->updatee);
p += sizeofW(StgUpdateFrame);
/* -----------------------------------------------------------------------------
- Lazy black holing.
-
- Whenever a thread returns to the scheduler after possibly doing
- some work, we have to run down the stack and black-hole all the
- closures referred to by update frames.
- -------------------------------------------------------------------------- */
-
-static void
-threadLazyBlackHole(StgTSO *tso)
-{
- StgClosure *frame;
- StgRetInfoTable *info;
- StgClosure *bh;
- StgPtr stack_end;
-
- stack_end = &tso->stack[tso->stack_size];
-
- frame = (StgClosure *)tso->sp;
-
- while (1) {
- info = get_ret_itbl(frame);
-
- switch (info->i.type) {
-
- case UPDATE_FRAME:
- bh = ((StgUpdateFrame *)frame)->updatee;
-
- /* if the thunk is already blackholed, it means we've also
- * already blackholed the rest of the thunks on this stack,
- * so we can stop early.
- *
- * The blackhole made for a CAF is a CAF_BLACKHOLE, so they
- * don't interfere with this optimisation.
- */
- if (bh->header.info == &stg_BLACKHOLE_info) {
- return;
- }
-
- if (bh->header.info != &stg_CAF_BLACKHOLE_info) {
-#if (!defined(LAZY_BLACKHOLING)) && defined(DEBUG)
- debugBelch("Unexpected lazy BHing required at 0x%04lx\n",(long)bh);
-#endif
-#ifdef PROFILING
- // @LDV profiling
- // We pretend that bh is now dead.
- LDV_recordDead_FILL_SLOP_DYNAMIC((StgClosure *)bh);
-#endif
- SET_INFO(bh,&stg_BLACKHOLE_info);
-
- // We pretend that bh has just been created.
- LDV_RECORD_CREATE(bh);
- }
-
- frame = (StgClosure *) ((StgUpdateFrame *)frame + 1);
- break;
-
- case STOP_FRAME:
- return;
-
- // normal stack frames; do nothing except advance the pointer
- default:
- frame = (StgClosure *)((StgPtr)frame + stack_frame_sizeW(frame));
- }
- }
-}
-
-
-/* -----------------------------------------------------------------------------
* Stack squeezing
*
* Code largely pinched from old RTS, then hacked to bits. We also do
struct stack_gap { StgWord gap_size; struct stack_gap *next_gap; };
static void
-threadSqueezeStack(StgTSO *tso)
+stackSqueeze(StgTSO *tso, StgPtr bottom)
{
StgPtr frame;
rtsBool prev_was_update_frame;
StgClosure *updatee = NULL;
- StgPtr bottom;
StgRetInfoTable *info;
StgWord current_gap_size;
struct stack_gap *gap;
// contains two values: the size of the gap, and the distance
// to the next gap (or the stack top).
- bottom = &(tso->stack[tso->stack_size]);
-
frame = tso->sp;
ASSERT(frame < bottom);
{
StgUpdateFrame *upd = (StgUpdateFrame *)frame;
- if (upd->updatee->header.info == &stg_BLACKHOLE_info) {
-
- // found a BLACKHOLE'd update frame; we've been here
- // before, in a previous GC, so just break out.
-
- // Mark the end of the gap, if we're in one.
- if (current_gap_size != 0) {
- gap = (struct stack_gap *)(frame-sizeofW(StgUpdateFrame));
- }
-
- frame += sizeofW(StgUpdateFrame);
- goto done_traversing;
- }
-
if (prev_was_update_frame) {
TICK_UPD_SQUEEZED();
// single update frame, or the topmost update frame in a series
else {
- StgClosure *bh = upd->updatee;
-
- // Do lazy black-holing
- if (bh->header.info != &stg_BLACKHOLE_info &&
- bh->header.info != &stg_CAF_BLACKHOLE_info) {
-#if (!defined(LAZY_BLACKHOLING)) && defined(DEBUG)
- debugBelch("Unexpected lazy BHing required at 0x%04lx",(long)bh);
-#endif
-#ifdef DEBUG
- // zero out the slop so that the sanity checker can tell
- // where the next closure is.
- DEBUG_FILL_SLOP(bh);
-#endif
-#ifdef PROFILING
- // We pretend that bh is now dead.
- // ToDo: is the slop filling the same as DEBUG_FILL_SLOP?
- LDV_recordDead_FILL_SLOP_DYNAMIC((StgClosure *)bh);
-#endif
- // Todo: maybe use SET_HDR() and remove LDV_RECORD_CREATE()?
- SET_INFO(bh,&stg_BLACKHOLE_info);
-
- // We pretend that bh has just been created.
- LDV_RECORD_CREATE(bh);
- }
-
prev_was_update_frame = rtsTrue;
updatee = upd->updatee;
frame += sizeofW(StgUpdateFrame);
}
}
-done_traversing:
-
+ if (current_gap_size != 0) {
+ gap = (struct stack_gap *) (frame - sizeofW(StgUpdateFrame));
+ }
+
// Now we have a stack with gaps in it, and we have to walk down
// shoving the stack up to fill in the gaps. A diagram might
// help:
* turned on.
* -------------------------------------------------------------------------- */
void
-threadPaused(StgTSO *tso)
+threadPaused(Capability *cap, StgTSO *tso)
{
- if ( RtsFlags.GcFlags.squeezeUpdFrames == rtsTrue )
- threadSqueezeStack(tso); // does black holing too
- else
- threadLazyBlackHole(tso);
+ StgClosure *frame;
+ StgRetInfoTable *info;
+ StgClosure *bh;
+ StgPtr stack_end;
+ nat words_to_squeeze = 0;
+ nat weight = 0;
+ nat weight_pending = 0;
+ rtsBool prev_was_update_frame;
+
+ stack_end = &tso->stack[tso->stack_size];
+
+ frame = (StgClosure *)tso->sp;
+
+ while (1) {
+ // If we've already marked this frame, then stop here.
+ if (frame->header.info == (StgInfoTable *)&stg_marked_upd_frame_info) {
+ goto end;
+ }
+
+ info = get_ret_itbl(frame);
+
+ switch (info->i.type) {
+
+ case UPDATE_FRAME:
+
+ SET_INFO(frame, (StgInfoTable *)&stg_marked_upd_frame_info);
+
+ bh = ((StgUpdateFrame *)frame)->updatee;
+
+ if (closure_IND(bh) || bh->header.info == &stg_BLACKHOLE_info) {
+ IF_DEBUG(squeeze, debugBelch("suspending duplicate work: %ld words of stack\n", (StgPtr)frame - tso->sp));
+
+ // If this closure is already an indirection, then
+ // suspend the computation up to this point:
+ suspendComputation(cap,tso,(StgPtr)frame);
+
+ // Now drop the update frame, and arrange to return
+ // the value to the frame underneath:
+ tso->sp = (StgPtr)frame + sizeofW(StgUpdateFrame) - 2;
+ tso->sp[1] = (StgWord)bh;
+ tso->sp[0] = (W_)&stg_enter_info;
+
+ // And continue with threadPaused; there might be
+ // yet more computation to suspend.
+ threadPaused(cap,tso);
+ return;
+ }
+
+ if (bh->header.info != &stg_CAF_BLACKHOLE_info) {
+#if (!defined(LAZY_BLACKHOLING)) && defined(DEBUG)
+ debugBelch("Unexpected lazy BHing required at 0x%04lx\n",(long)bh);
+#endif
+ // zero out the slop so that the sanity checker can tell
+ // where the next closure is.
+ DEBUG_FILL_SLOP(bh);
+#ifdef PROFILING
+ // @LDV profiling
+ // We pretend that bh is now dead.
+ LDV_recordDead_FILL_SLOP_DYNAMIC((StgClosure *)bh);
+#endif
+ SET_INFO(bh,&stg_BLACKHOLE_info);
+
+ // We pretend that bh has just been created.
+ LDV_RECORD_CREATE(bh);
+ }
+
+ frame = (StgClosure *) ((StgUpdateFrame *)frame + 1);
+ if (prev_was_update_frame) {
+ words_to_squeeze += sizeofW(StgUpdateFrame);
+ weight += weight_pending;
+ weight_pending = 0;
+ }
+ prev_was_update_frame = rtsTrue;
+ break;
+
+ case STOP_FRAME:
+ goto end;
+
+ // normal stack frames; do nothing except advance the pointer
+ default:
+ {
+ nat frame_size = stack_frame_sizeW(frame);
+ weight_pending += frame_size;
+ frame = (StgClosure *)((StgPtr)frame + frame_size);
+ prev_was_update_frame = rtsFalse;
+ }
+ }
+ }
+
+end:
+ IF_DEBUG(squeeze,
+ debugBelch("words_to_squeeze: %d, weight: %d, squeeze: %s\n",
+ words_to_squeeze, weight,
+ weight < words_to_squeeze ? "YES" : "NO"));
+
+ // Should we squeeze or not? Arbitrary heuristic: we squeeze if
+ // the number of words we have to shift down is less than the
+ // number of stack words we squeeze away by doing so.
+ if (RtsFlags.GcFlags.squeezeUpdFrames == rtsTrue &&
+ weight < words_to_squeeze) {
+ stackSqueeze(tso, (StgPtr)frame);
+ }
}
/* -----------------------------------------------------------------------------
projects / ghc-hetmet.git / blobdiff