/* Used to avoid long recursion due to selector thunks
*/
-lnat thunk_selector_depth = 0;
#define MAX_THUNK_SELECTOR_DEPTH 16
static StgClosure * eval_thunk_selector (StgSelector * p, rtsBool);
-STATIC_INLINE void
-upd_evacuee(StgClosure *p, StgClosure *dest)
+STATIC_INLINE StgPtr
+alloc_for_copy (nat size, step *stp)
{
- // not true: (ToDo: perhaps it should be)
- // ASSERT(Bdescr((P_)dest)->flags & BF_EVACUATED);
- SET_INFO(p, &stg_EVACUATED_info);
- ((StgEvacuated *)p)->evacuee = dest;
+ StgPtr to;
+ step_workspace *ws;
+ bdescr *bd;
+
+ /* Find out where we're going, using the handy "to" pointer in
+ * the step of the source object. If it turns out we need to
+ * evacuate to an older generation, adjust it here (see comment
+ * by evacuate()).
+ */
+ if (stp->gen_no < gct->evac_gen) {
+ if (gct->eager_promotion) {
+ stp = &generations[gct->evac_gen].steps[0];
+ } else {
+ gct->failed_to_evac = rtsTrue;
+ }
+ }
+
+ ws = &gct->steps[stp->gen_no][stp->no];
+
+ /* chain a new block onto the to-space for the destination step if
+ * necessary.
+ */
+ bd = ws->todo_bd;
+ to = bd->free;
+ if (to + size >= bd->start + BLOCK_SIZE_W) {
+ bd = gc_alloc_todo_block(ws);
+ to = bd->free;
+ }
+ bd->free = to + size;
+
+ return to;
}
+
+STATIC_INLINE StgPtr
+alloc_for_copy_noscav (nat size, step *stp)
+{
+ StgPtr to;
+ step_workspace *ws;
+ bdescr *bd;
+ /* Find out where we're going, using the handy "to" pointer in
+ * the step of the source object. If it turns out we need to
+ * evacuate to an older generation, adjust it here (see comment
+ * by evacuate()).
+ */
+ if (stp->gen_no < gct->evac_gen) {
+ if (gct->eager_promotion) {
+ stp = &generations[gct->evac_gen].steps[0];
+ } else {
+ gct->failed_to_evac = rtsTrue;
+ }
+ }
+
+ ws = &gct->steps[stp->gen_no][stp->no];
+
+ /* chain a new block onto the to-space for the destination step if
+ * necessary.
+ */
+ bd = ws->scavd_list;
+ to = bd->free;
+ if (to + size >= bd->start + BLOCK_SIZE_W) {
+ bd = gc_alloc_scavd_block(ws);
+ to = bd->free;
+ }
+ bd->free = to + size;
+ return to;
+}
+
STATIC_INLINE StgClosure *
copy_tag(StgClosure *src, nat size, step *stp,StgWord tag)
{
StgPtr to, from;
nat i;
-#ifdef PROFILING
- // @LDV profiling
- nat size_org = size;
+ StgWord info;
+
+#ifdef THREADED_RTS
+ do {
+ info = xchg((StgPtr)&src->header.info, (W_)&stg_WHITEHOLE_info);
+ // so.. what is it?
+ } while (info == (W_)&stg_WHITEHOLE_info);
+ if (info == (W_)&stg_EVACUATED_info) {
+ src->header.info = (const StgInfoTable *)info;
+ return evacuate(src); // does the failed_to_evac stuff
+ }
+#else
+ info = (W_)src->header.info;
+ src->header.info = &stg_EVACUATED_info;
#endif
- TICK_GC_WORDS_COPIED(size);
- /* Find out where we're going, using the handy "to" pointer in
- * the step of the source object. If it turns out we need to
- * evacuate to an older generation, adjust it here (see comment
- * by evacuate()).
- */
- if (stp->gen_no < evac_gen) {
- 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
- * necessary.
- */
- if (stp->hp + size >= stp->hpLim) {
- gc_alloc_block(stp);
- }
+ to = alloc_for_copy(size,stp);
+
+ TICK_GC_WORDS_COPIED(size);
- to = stp->hp;
- from = (StgPtr)src;
- stp->hp = to + size;
- for (i = 0; i < size; i++) { // unroll for small i
- to[i] = from[i];
- }
+ from = (StgPtr)src;
+ to[0] = info;
+ for (i = 1; i < size; i++) { // unroll for small i
+ to[i] = from[i];
+ }
+
+ // retag pointer before updating EVACUATE closure and returning
+ to = (StgPtr)TAG_CLOSURE(tag,(StgClosure*)to);
- /* retag pointer before updating EVACUATE closure and returning */
- to = (StgPtr)TAG_CLOSURE(tag,(StgClosure*)to);
+// if (to+size+2 < bd->start + BLOCK_SIZE_W) {
+// __builtin_prefetch(to + size + 2, 1);
+// }
- upd_evacuee((StgClosure *)from,(StgClosure *)to);
+ ((StgEvacuated*)from)->evacuee = (StgClosure *)to;
+#ifdef THREADED_RTS
+ write_barrier();
+ ((StgEvacuated*)from)->header.info = &stg_EVACUATED_info;
+#endif
#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(from, size_org);
+ // 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(from, size);
#endif
- return (StgClosure *)to;
+ return (StgClosure *)to;
}
+
// Same as copy() above, except the object will be allocated in memory
// that will not be scavenged. Used for object that have no pointer
STATIC_INLINE StgClosure *
copy_noscav_tag(StgClosure *src, nat size, step *stp, StgWord tag)
{
- StgPtr to, from;
- nat i;
-#ifdef PROFILING
- // @LDV profiling
- nat size_org = size;
+ StgPtr to, from;
+ nat i;
+ StgWord info;
+
+#ifdef THREADED_RTS
+ do {
+ info = xchg((StgPtr)&src->header.info, (W_)&stg_WHITEHOLE_info);
+ } while (info == (W_)&stg_WHITEHOLE_info);
+ if (info == (W_)&stg_EVACUATED_info) {
+ src->header.info = (const StgInfoTable *)info;
+ return evacuate(src); // does the failed_to_evac stuff
+ }
+#else
+ info = (W_)src->header.info;
+ src->header.info = &stg_EVACUATED_info;
#endif
+
+ to = alloc_for_copy_noscav(size,stp);
- TICK_GC_WORDS_COPIED(size);
- /* Find out where we're going, using the handy "to" pointer in
- * the step of the source object. If it turns out we need to
- * evacuate to an older generation, adjust it here (see comment
- * by evacuate()).
- */
- if (stp->gen_no < evac_gen) {
- 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
- * necessary.
- */
- if (stp->scavd_hp + size >= stp->scavd_hpLim) {
- gc_alloc_scavd_block(stp);
- }
-
- to = stp->scavd_hp;
- from = (StgPtr)src;
- stp->scavd_hp = to + size;
- for (i = 0; i < size; i++) { // unroll for small i
- to[i] = from[i];
- }
-
- /* retag pointer before updating EVACUATE closure and returning */
- to = (StgPtr)TAG_CLOSURE(tag,(StgClosure*)to);
+ TICK_GC_WORDS_COPIED(size);
+
+ from = (StgPtr)src;
+ to[0] = info;
+ for (i = 1; i < size; i++) { // unroll for small i
+ to[i] = from[i];
+ }
- upd_evacuee((StgClosure *)from,(StgClosure *)to);
+ // retag pointer before updating EVACUATE closure and returning
+ to = (StgPtr)TAG_CLOSURE(tag,(StgClosure*)to);
+ ((StgEvacuated*)from)->evacuee = (StgClosure *)to;
+#ifdef THREADED_RTS
+ write_barrier();
+ ((StgEvacuated*)from)->header.info = &stg_EVACUATED_info;
+#endif
+
#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(from, size_org);
+ // 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(from, size);
#endif
- return (StgClosure *)to;
+ return (StgClosure *)to;
}
+
/* Special version of copy() for when we only want to copy the info
* pointer of an object, but reserve some padding after it. This is
* used to optimise evacuation of BLACKHOLEs.
*/
-
-
static StgClosure *
copyPart(StgClosure *src, nat size_to_reserve, nat size_to_copy, step *stp)
{
- P_ dest, to, from;
-#ifdef PROFILING
- // @LDV profiling
- nat size_to_copy_org = size_to_copy;
+ StgPtr to, from;
+ nat i;
+ StgWord info;
+
+#ifdef THREADED_RTS
+ do {
+ info = xchg((StgPtr)&src->header.info, (W_)&stg_WHITEHOLE_info);
+ } while (info == (W_)&stg_WHITEHOLE_info);
+ if (info == (W_)&stg_EVACUATED_info) {
+ src->header.info = (const StgInfoTable *)info;
+ return evacuate(src); // does the failed_to_evac stuff
+ }
+#else
+ info = (W_)src->header.info;
+ src->header.info = &stg_EVACUATED_info;
#endif
+
+ to = alloc_for_copy(size_to_reserve, stp);
- TICK_GC_WORDS_COPIED(size_to_copy);
- if (stp->gen_no < evac_gen) {
- if (eager_promotion) {
- stp = &generations[evac_gen].steps[0];
- } else {
- failed_to_evac = rtsTrue;
- }
- }
-
- if (stp->hp + size_to_reserve >= stp->hpLim) {
- gc_alloc_block(stp);
- }
+ TICK_GC_WORDS_COPIED(size_to_copy);
- for(to = stp->hp, from = (P_)src; size_to_copy>0; --size_to_copy) {
- *to++ = *from++;
- }
-
- dest = stp->hp;
- stp->hp += size_to_reserve;
- upd_evacuee(src,(StgClosure *)dest);
+ from = (StgPtr)src;
+ to[0] = info;
+ for (i = 1; i < size_to_copy; i++) { // unroll for small i
+ to[i] = from[i];
+ }
+
+ ((StgEvacuated*)from)->evacuee = (StgClosure *)to;
+#ifdef THREADED_RTS
+ write_barrier();
+ ((StgEvacuated*)from)->header.info = &stg_EVACUATED_info;
+#endif
+
#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.
- // size_to_copy_org is wrong because the closure already occupies size_to_reserve
- // words.
- SET_EVACUAEE_FOR_LDV(src, size_to_reserve);
- // fill the slop
- if (size_to_reserve - size_to_copy_org > 0)
- LDV_FILL_SLOP(stp->hp - 1, (int)(size_to_reserve - size_to_copy_org));
+ // 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(from, size_to_reserve);
+ // fill the slop
+ if (size_to_reserve - size_to_copy > 0)
+ LDV_FILL_SLOP(to + size_to_copy - 1, (int)(size_to_reserve - size_to_copy));
#endif
- return (StgClosure *)dest;
+ return (StgClosure *)to;
}
{
bdescr *bd = Bdescr(p);
step *stp;
+ step_workspace *ws;
// object must be at the beginning of the block (or be a ByteArray)
ASSERT(get_itbl((StgClosure *)p)->type == ARR_WORDS ||
// already evacuated?
if (bd->flags & BF_EVACUATED) {
- /* Don't forget to set the failed_to_evac flag if we didn't get
+ /* Don't forget to set the gct->failed_to_evac flag if we didn't get
* the desired destination (see comments in evacuate()).
*/
- if (bd->gen_no < evac_gen) {
- failed_to_evac = rtsTrue;
+ if (bd->gen_no < gct->evac_gen) {
+ gct->failed_to_evac = rtsTrue;
TICK_GC_FAILED_PROMOTION();
}
return;
}
stp = bd->step;
+
+ ACQUIRE_SPIN_LOCK(&stp->sync_large_objects);
// remove from large_object list
if (bd->u.back) {
bd->u.back->link = bd->link;
if (bd->link) {
bd->link->u.back = bd->u.back;
}
+ RELEASE_SPIN_LOCK(&stp->sync_large_objects);
/* link it on to the evacuated large object list of the destination step
*/
stp = bd->step->to;
- if (stp->gen_no < evac_gen) {
- if (eager_promotion) {
- stp = &generations[evac_gen].steps[0];
+ if (stp->gen_no < gct->evac_gen) {
+ if (gct->eager_promotion) {
+ stp = &generations[gct->evac_gen].steps[0];
} else {
- failed_to_evac = rtsTrue;
+ gct->failed_to_evac = rtsTrue;
}
}
+ ws = &gct->steps[stp->gen_no][stp->no];
bd->step = stp;
bd->gen_no = stp->gen_no;
- bd->link = stp->new_large_objects;
- stp->new_large_objects = bd;
+ bd->link = ws->todo_large_objects;
+ ws->todo_large_objects = bd;
bd->flags |= BF_EVACUATED;
}
This is called (eventually) for every live object in the system.
The caller to evacuate specifies a desired generation in the
- evac_gen global variable. The following conditions apply to
+ gct->evac_gen thread-lock variable. The following conditions apply to
evacuating an object which resides in generation M when we're
collecting up to generation N
- if M >= evac_gen
+ if M >= gct->evac_gen
if M > N do nothing
else evac to step->to
- if M < evac_gen evac to evac_gen, step 0
+ if M < gct->evac_gen evac to gct->evac_gen, step 0
if the object is already evacuated, then we check which generation
it now resides in.
- if M >= evac_gen do nothing
- if M < evac_gen set failed_to_evac flag to indicate that we
- didn't manage to evacuate this object into evac_gen.
+ if M >= gct->evac_gen do nothing
+ if M < gct->evac_gen set gct->failed_to_evac flag to indicate that we
+ didn't manage to evacuate this object into gct->evac_gen.
OPTIMISATION NOTES:
switch (info->type) {
case THUNK_STATIC:
- if (info->srt_bitmap != 0 &&
- *THUNK_STATIC_LINK((StgClosure *)q) == NULL) {
- *THUNK_STATIC_LINK((StgClosure *)q) = static_objects;
- static_objects = (StgClosure *)q;
+ if (info->srt_bitmap != 0) {
+ ACQUIRE_SPIN_LOCK(&static_objects_sync);
+ if (*THUNK_STATIC_LINK((StgClosure *)q) == NULL) {
+ *THUNK_STATIC_LINK((StgClosure *)q) = static_objects;
+ static_objects = (StgClosure *)q;
+ }
+ RELEASE_SPIN_LOCK(&static_objects_sync);
}
return q;
case FUN_STATIC:
- if (info->srt_bitmap != 0 &&
- *FUN_STATIC_LINK((StgClosure *)q) == NULL) {
- *FUN_STATIC_LINK((StgClosure *)q) = static_objects;
- static_objects = (StgClosure *)q;
+ if (info->srt_bitmap != 0) {
+ ACQUIRE_SPIN_LOCK(&static_objects_sync);
+ if (*FUN_STATIC_LINK((StgClosure *)q) == NULL) {
+ *FUN_STATIC_LINK((StgClosure *)q) = static_objects;
+ static_objects = (StgClosure *)q;
+ }
+ RELEASE_SPIN_LOCK(&static_objects_sync);
}
return q;
* on the CAF list, so don't do anything with it here (we'll
* scavenge it later).
*/
- if (((StgIndStatic *)q)->saved_info == NULL
- && *IND_STATIC_LINK((StgClosure *)q) == NULL) {
- *IND_STATIC_LINK((StgClosure *)q) = static_objects;
- static_objects = (StgClosure *)q;
+ if (((StgIndStatic *)q)->saved_info == NULL) {
+ ACQUIRE_SPIN_LOCK(&static_objects_sync);
+ if (*IND_STATIC_LINK((StgClosure *)q) == NULL) {
+ *IND_STATIC_LINK((StgClosure *)q) = static_objects;
+ static_objects = (StgClosure *)q;
+ }
+ RELEASE_SPIN_LOCK(&static_objects_sync);
}
return q;
case CONSTR_STATIC:
if (*STATIC_LINK(info,(StgClosure *)q) == NULL) {
- *STATIC_LINK(info,(StgClosure *)q) = static_objects;
- static_objects = (StgClosure *)q;
+ ACQUIRE_SPIN_LOCK(&static_objects_sync);
+ // re-test, after acquiring lock
+ if (*STATIC_LINK(info,(StgClosure *)q) == NULL) {
+ *STATIC_LINK(info,(StgClosure *)q) = static_objects;
+ static_objects = (StgClosure *)q;
+ }
+ RELEASE_SPIN_LOCK(&static_objects_sync);
/* I am assuming that static_objects pointers are not
* written to other objects, and thus, no need to retag. */
}
if (bd->gen_no > N) {
/* Can't evacuate this object, because it's in a generation
* older than the ones we're collecting. Let's hope that it's
- * in evac_gen or older, or we will have to arrange to track
+ * in gct->evac_gen or older, or we will have to arrange to track
* this pointer using the mutable list.
*/
- if (bd->gen_no < evac_gen) {
+ if (bd->gen_no < gct->evac_gen) {
// nope
- failed_to_evac = rtsTrue;
+ gct->failed_to_evac = rtsTrue;
TICK_GC_FAILED_PROMOTION();
}
return TAG_CLOSURE(tag,q);
* object twice, for example).
*/
if (bd->flags & BF_EVACUATED) {
- if (bd->gen_no < evac_gen) {
- failed_to_evac = rtsTrue;
+ if (bd->gen_no < gct->evac_gen) {
+ gct->failed_to_evac = rtsTrue;
TICK_GC_FAILED_PROMOTION();
}
return TAG_CLOSURE(tag,q);
switch (info->type) {
+ case WHITEHOLE:
+ goto loop;
+
case MUT_VAR_CLEAN:
case MUT_VAR_DIRTY:
case MVAR_CLEAN:
case EVACUATED:
/* Already evacuated, just return the forwarding address.
- * HOWEVER: if the requested destination generation (evac_gen) is
+ * HOWEVER: if the requested destination generation (gct->evac_gen) is
* older than the actual generation (because the object was
* already evacuated to a younger generation) then we have to
- * set the failed_to_evac flag to indicate that we couldn't
+ * set the gct->failed_to_evac flag to indicate that we couldn't
* manage to promote the object to the desired generation.
*/
/*
* current object would be evacuated to, so we only do the full
* check if stp is too low.
*/
- if (evac_gen > 0 && stp->gen_no < evac_gen) { // optimisation
+ if (gct->evac_gen > 0 && stp->gen_no < gct->evac_gen) { // optimisation
StgClosure *p = ((StgEvacuated*)q)->evacuee;
- if (HEAP_ALLOCED(p) && Bdescr((P_)p)->gen_no < evac_gen) {
- failed_to_evac = rtsTrue;
+ if (HEAP_ALLOCED(p) && Bdescr((P_)p)->gen_no < gct->evac_gen) {
+ gct->failed_to_evac = rtsTrue;
TICK_GC_FAILED_PROMOTION();
}
}
// recursively evaluate this selector. We don't want to
// recurse indefinitely, so we impose a depth bound.
- if (thunk_selector_depth >= MAX_THUNK_SELECTOR_DEPTH) {
+ if (gct->thunk_selector_depth >= MAX_THUNK_SELECTOR_DEPTH) {
goto bale_out;
}
- thunk_selector_depth++;
+ gct->thunk_selector_depth++;
// rtsFalse says "don't evacuate the result". It will,
// however, update any THUNK_SELECTORs that are evaluated
// along the way.
val = eval_thunk_selector((StgSelector *)selectee, rtsFalse);
- thunk_selector_depth--;
+ gct->thunk_selector_depth--;
// did we actually manage to evaluate it?
if (val == selectee) goto bale_out;