// stats. Every mutable list is copied during every GC.
if (g > 0) {
for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
- copied += (bd->free - bd->start) * sizeof(StgWord);
+ copied += bd->free - bd->start;
}
}
resetNurseries();
// start any pending finalizers
+ RELEASE_SM_LOCK;
scheduleFinalizers(last_free_capability, old_weak_ptr_list);
+ ACQUIRE_SM_LOCK;
// send exceptions to any threads which were about to die
resurrectThreads(resurrected_threads);
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;
}
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);
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;
}
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;
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;
}
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;
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 (1 /*RtsFlags.GcFlags.squeezeUpdFrames == rtsTrue &&
+ weight < words_to_squeeze*/) {
+ stackSqueeze(tso, (StgPtr)frame);
+ }
}
/* -----------------------------------------------------------------------------