static void
checkPAP (StgClosure *tagged_fun, StgClosure** payload, StgWord n_args)
{
StgClosure *fun;
StgFunInfoTable *fun_info;
fun = UNTAG_CLOSURE(tagged_fun);
ASSERT(LOOKS_LIKE_CLOSURE_PTR(fun));
fun_info = get_fun_itbl(fun);
switch (fun_info->f.fun_type) {
case ARG_GEN:
checkSmallBitmap( (StgPtr)payload,
BITMAP_BITS(fun_info->f.b.bitmap), n_args );
break;
case ARG_GEN_BIG:
checkLargeBitmap( (StgPtr)payload,
GET_FUN_LARGE_BITMAP(fun_info),
n_args );
break;
case ARG_BCO:
checkLargeBitmap( (StgPtr)payload,
BCO_BITMAP(fun),
n_args );
break;
default:
checkSmallBitmap( (StgPtr)payload,
BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]),
n_args );
break;
}
ASSERT(fun_info->f.arity > TAG_MASK ? GET_CLOSURE_TAG(tagged_fun) == 0
: GET_CLOSURE_TAG(tagged_fun) == fun_info->f.arity);
}
/*
* get at the real stuff...remove indirections.
*/
static StgClosure*
removeIndirections (StgClosure* p)
{
StgClosure* q;
while (1)
{
q = UNTAG_CLOSURE(p);
switch (get_itbl(q)->type) {
case IND:
case IND_STATIC:
p = ((StgInd *)q)->indirectee;
continue;
case BLACKHOLE:
p = ((StgInd *)q)->indirectee;
if (GET_CLOSURE_TAG(p) != 0) {
continue;
} else {
break;
}
default:
break;
}
return p;
}
}
// Traverse a threaded chain and pull out the info pointer at the end.
// The info pointer is also tagged with the appropriate pointer tag
// for this closure, which should be attached to the pointer
// subsequently passed to unthread().
STATIC_INLINE StgWord
get_threaded_info( StgPtr p )
{
StgWord q;
q = (W_)GET_INFO(UNTAG_CLOSURE((StgClosure *)p));
loop:
switch (GET_CLOSURE_TAG((StgClosure *)q))
{
case 0:
ASSERT(LOOKS_LIKE_INFO_PTR(q));
return q;
case 1:
{
StgWord r = *(StgPtr)(q-1);
ASSERT(LOOKS_LIKE_INFO_PTR(UNTAG_CLOSURE((StgClosure *)r)));
return r;
}
case 2:
q = *(StgPtr)(q-2);
goto loop;
default:
barf("get_threaded_info");
}
}
STATIC_INLINE void
unthread( StgPtr p, StgWord free )
{
StgWord q, r;
StgPtr q0;
q = *p;
loop:
switch (GET_CLOSURE_TAG((StgClosure *)q))
{
case 0:
// nothing to do; the chain is length zero
return;
case 1:
q0 = (StgPtr)(q-1);
r = *q0; // r is the info ptr, tagged with the pointer-tag
*q0 = free;
*p = (StgWord)UNTAG_CLOSURE((StgClosure *)r);
return;
case 2:
q0 = (StgPtr)(q-2);
r = *q0;
*q0 = free;
q = r;
goto loop;
default:
barf("unthread");
}
}
/*
* get at the real stuff...remove indirections.
* It untags pointers before dereferencing and
* retags the real stuff with its tag (if there
* is any) when returning.
*
* ToDo: move to a better home.
*/
static
StgClosure*
removeIndirections(StgClosure* p)
{
StgWord tag = GET_CLOSURE_TAG(p);
StgClosure* q = UNTAG_CLOSURE(p);
while (get_itbl(q)->type == IND ||
get_itbl(q)->type == IND_STATIC ||
get_itbl(q)->type == IND_PERM) {
q = ((StgInd *)q)->indirectee;
tag = GET_CLOSURE_TAG(q);
q = UNTAG_CLOSURE(q);
}
return TAG_CLOSURE(tag,q);
}
void
evacuate_BLACKHOLE(StgClosure **p)
{
bdescr *bd;
uint32_t gen_no;
StgClosure *q;
const StgInfoTable *info;
q = *p;
// closure is required to be a heap-allocated BLACKHOLE
ASSERT(HEAP_ALLOCED_GC(q));
ASSERT(GET_CLOSURE_TAG(q) == 0);
bd = Bdescr((P_)q);
// blackholes can't be in a compact
ASSERT((bd->flags & BF_COMPACT) == 0);
// blackholes *can* be in a large object: when raiseAsync() creates an
// AP_STACK the payload might be large enough to create a large object.
// See #14497.
if (bd->flags & BF_LARGE) {
evacuate_large((P_)q);
return;
}
if (bd->flags & BF_EVACUATED) {
if (bd->gen_no < gct->evac_gen_no) {
gct->failed_to_evac = true;
TICK_GC_FAILED_PROMOTION();
}
return;
}
if (bd->flags & BF_MARKED) {
if (!is_marked((P_)q,bd)) {
mark((P_)q,bd);
push_mark_stack((P_)q);
}
return;
}
gen_no = bd->dest_no;
info = q->header.info;
if (IS_FORWARDING_PTR(info))
{
StgClosure *e = (StgClosure*)UN_FORWARDING_PTR(info);
*p = e;
if (gen_no < gct->evac_gen_no) { // optimisation
if (Bdescr((P_)e)->gen_no < gct->evac_gen_no) {
gct->failed_to_evac = true;
TICK_GC_FAILED_PROMOTION();
}
}
return;
}
ASSERT(INFO_PTR_TO_STRUCT(info)->type == BLACKHOLE);
copy(p,info,q,sizeofW(StgInd),gen_no);
}
STATIC_INLINE void
thread (StgClosure **p)
{
StgClosure *q0;
StgPtr q;
StgWord iptr;
bdescr *bd;
q0 = *p;
q = (StgPtr)UNTAG_CLOSURE(q0);
// It doesn't look like a closure at the moment, because the info
// ptr is possibly threaded:
// ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
if (HEAP_ALLOCED(q)) {
bd = Bdescr(q);
// a handy way to discover whether the ptr is into the
// compacted area of the old gen, is that the EVACUATED flag
// is zero (it's non-zero for all the other areas of live
// memory).
if ((bd->flags & BF_EVACUATED) == 0)
{
iptr = *q;
switch (GET_CLOSURE_TAG((StgClosure *)iptr))
{
case 0:
// this is the info pointer; we are creating a new chain.
// save the original tag at the end of the chain.
*p = (StgClosure *)((StgWord)iptr + GET_CLOSURE_TAG(q0));
*q = (StgWord)p + 1;
break;
case 1:
case 2:
// this is a chain of length 1 or more
*p = (StgClosure *)iptr;
*q = (StgWord)p + 2;
break;
}
}
}
}
STATIC_INLINE void
thread (StgClosure **p)
{
StgClosure *q0;
StgPtr q;
StgWord iptr;
bdescr *bd;
q0 = *p;
q = (StgPtr)UNTAG_CLOSURE(q0);
// It doesn't look like a closure at the moment, because the info
// ptr is possibly threaded:
// ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
if (HEAP_ALLOCED(q)) {
bd = Bdescr(q);
if (bd->flags & BF_MARKED)
{
iptr = *q;
switch (GET_CLOSURE_TAG((StgClosure *)iptr))
{
case 0:
// this is the info pointer; we are creating a new chain.
// save the original tag at the end of the chain.
*p = (StgClosure *)((StgWord)iptr + GET_CLOSURE_TAG(q0));
*q = (StgWord)p + 1;
break;
case 1:
case 2:
// this is a chain of length 1 or more
*p = (StgClosure *)iptr;
*q = (StgWord)p + 2;
break;
}
}
}
}
static void
eval_thunk_selector (StgClosure **q, StgSelector * p, rtsBool evac)
// NB. for legacy reasons, p & q are swapped around :(
{
nat field;
StgInfoTable *info;
StgWord info_ptr;
StgClosure *selectee;
StgSelector *prev_thunk_selector;
bdescr *bd;
StgClosure *val;
prev_thunk_selector = NULL;
// this is a chain of THUNK_SELECTORs that we are going to update
// to point to the value of the current THUNK_SELECTOR. Each
// closure on the chain is a WHITEHOLE, and points to the next in the
// chain with payload[0].
selector_chain:
bd = Bdescr((StgPtr)p);
if (HEAP_ALLOCED_GC(p)) {
// If the THUNK_SELECTOR is in to-space or in a generation that we
// are not collecting, then bale out early. We won't be able to
// save any space in any case, and updating with an indirection is
// trickier in a non-collected gen: we would have to update the
// mutable list.
if (bd->flags & BF_EVACUATED) {
unchain_thunk_selectors(prev_thunk_selector, (StgClosure *)p);
*q = (StgClosure *)p;
// shortcut, behave as for: if (evac) evacuate(q);
if (evac && bd->gen_no < gct->evac_gen_no) {
gct->failed_to_evac = rtsTrue;
TICK_GC_FAILED_PROMOTION();
}
return;
}
// we don't update THUNK_SELECTORS in the compacted
// generation, because compaction does not remove the INDs
// that result, this causes confusion later
// (scavenge_mark_stack doesn't deal with IND). BEWARE! This
// bit is very tricky to get right. If you make changes
// around here, test by compiling stage 3 with +RTS -c -RTS.
if (bd->flags & BF_MARKED) {
// must call evacuate() to mark this closure if evac==rtsTrue
*q = (StgClosure *)p;
if (evac) evacuate(q);
unchain_thunk_selectors(prev_thunk_selector, (StgClosure *)p);
return;
}
}
// WHITEHOLE the selector thunk, since it is now under evaluation.
// This is important to stop us going into an infinite loop if
// this selector thunk eventually refers to itself.
#if defined(THREADED_RTS)
// In threaded mode, we'll use WHITEHOLE to lock the selector
// thunk while we evaluate it.
{
do {
info_ptr = xchg((StgPtr)&p->header.info, (W_)&stg_WHITEHOLE_info);
} while (info_ptr == (W_)&stg_WHITEHOLE_info);
// make sure someone else didn't get here first...
if (IS_FORWARDING_PTR(info_ptr) ||
INFO_PTR_TO_STRUCT((StgInfoTable *)info_ptr)->type != THUNK_SELECTOR) {
// v. tricky now. The THUNK_SELECTOR has been evacuated
// by another thread, and is now either a forwarding ptr or IND.
// We need to extract ourselves from the current situation
// as cleanly as possible.
// - unlock the closure
// - update *q, we may have done *some* evaluation
// - if evac, we need to call evacuate(), because we
// need the write-barrier stuff.
// - undo the chain we've built to point to p.
SET_INFO((StgClosure *)p, (const StgInfoTable *)info_ptr);
*q = (StgClosure *)p;
if (evac) evacuate(q);
unchain_thunk_selectors(prev_thunk_selector, (StgClosure *)p);
return;
}
}
#else
// Save the real info pointer (NOTE: not the same as get_itbl()).
info_ptr = (StgWord)p->header.info;
SET_INFO((StgClosure *)p,&stg_WHITEHOLE_info);
#endif
field = INFO_PTR_TO_STRUCT((StgInfoTable *)info_ptr)->layout.selector_offset;
// The selectee might be a constructor closure,
// so we untag the pointer.
selectee = UNTAG_CLOSURE(p->selectee);
selector_loop:
// selectee now points to the closure that we're trying to select
// a field from. It may or may not be in to-space: we try not to
// end up in to-space, but it's impractical to avoid it in
// general. The compacting GC scatters to-space pointers in
// from-space during marking, for example. We rely on the property
// that evacuate() doesn't mind if it gets passed a to-space pointer.
info = (StgInfoTable*)selectee->header.info;
if (IS_FORWARDING_PTR(info)) {
// We don't follow pointers into to-space; the constructor
// has already been evacuated, so we won't save any space
// leaks by evaluating this selector thunk anyhow.
goto bale_out;
}
info = INFO_PTR_TO_STRUCT(info);
switch (info->type) {
case WHITEHOLE:
goto bale_out; // about to be evacuated by another thread (or a loop).
case CONSTR:
case CONSTR_1_0:
case CONSTR_0_1:
case CONSTR_2_0:
case CONSTR_1_1:
case CONSTR_0_2:
case CONSTR_STATIC:
case CONSTR_NOCAF_STATIC:
{
// check that the size is in range
ASSERT(field < (StgWord32)(info->layout.payload.ptrs +
info->layout.payload.nptrs));
// Select the right field from the constructor
val = selectee->payload[field];
#ifdef PROFILING
// For the purposes of LDV profiling, we have destroyed
// the original selector thunk, p.
if (era > 0) {
// Only modify the info pointer when LDV profiling is
// enabled. Note that this is incompatible with parallel GC,
// because it would allow other threads to start evaluating
// the same selector thunk.
SET_INFO((StgClosure*)p, (StgInfoTable *)info_ptr);
OVERWRITING_CLOSURE((StgClosure*)p);
SET_INFO((StgClosure*)p, &stg_WHITEHOLE_info);
}
#endif
// the closure in val is now the "value" of the
// THUNK_SELECTOR in p. However, val may itself be a
// THUNK_SELECTOR, in which case we want to continue
// evaluating until we find the real value, and then
// update the whole chain to point to the value.
val_loop:
info_ptr = (StgWord)UNTAG_CLOSURE(val)->header.info;
if (!IS_FORWARDING_PTR(info_ptr))
{
info = INFO_PTR_TO_STRUCT((StgInfoTable *)info_ptr);
switch (info->type) {
case IND:
case IND_STATIC:
val = ((StgInd *)val)->indirectee;
goto val_loop;
case THUNK_SELECTOR:
((StgClosure*)p)->payload[0] = (StgClosure *)prev_thunk_selector;
prev_thunk_selector = p;
p = (StgSelector*)val;
goto selector_chain;
default:
break;
}
}
((StgClosure*)p)->payload[0] = (StgClosure *)prev_thunk_selector;
prev_thunk_selector = p;
*q = val;
// update the other selectors in the chain *before*
// evacuating the value. This is necessary in the case
// where the value turns out to be one of the selectors
// in the chain (i.e. we have a loop), and evacuating it
// would corrupt the chain.
unchain_thunk_selectors(prev_thunk_selector, val);
// evacuate() cannot recurse through
// eval_thunk_selector(), because we know val is not
// a THUNK_SELECTOR.
if (evac) evacuate(q);
return;
}
case IND:
case IND_STATIC:
// Again, we might need to untag a constructor.
selectee = UNTAG_CLOSURE( ((StgInd *)selectee)->indirectee );
goto selector_loop;
case BLACKHOLE:
{
StgClosure *r;
const StgInfoTable *i;
r = ((StgInd*)selectee)->indirectee;
// establish whether this BH has been updated, and is now an
// indirection, as in evacuate().
if (GET_CLOSURE_TAG(r) == 0) {
i = r->header.info;
if (IS_FORWARDING_PTR(i)) {
r = (StgClosure *)UN_FORWARDING_PTR(i);
i = r->header.info;
}
if (i == &stg_TSO_info
|| i == &stg_WHITEHOLE_info
|| i == &stg_BLOCKING_QUEUE_CLEAN_info
|| i == &stg_BLOCKING_QUEUE_DIRTY_info) {
goto bale_out;
}
ASSERT(i != &stg_IND_info);
}
selectee = UNTAG_CLOSURE( ((StgInd *)selectee)->indirectee );
goto selector_loop;
}
case THUNK_SELECTOR:
{
StgClosure *val;
// recursively evaluate this selector. We don't want to
// recurse indefinitely, so we impose a depth bound.
if (gct->thunk_selector_depth >= MAX_THUNK_SELECTOR_DEPTH) {
goto bale_out;
}
gct->thunk_selector_depth++;
// rtsFalse says "don't evacuate the result". It will,
// however, update any THUNK_SELECTORs that are evaluated
// along the way.
eval_thunk_selector(&val, (StgSelector*)selectee, rtsFalse);
gct->thunk_selector_depth--;
// did we actually manage to evaluate it?
if (val == selectee) goto bale_out;
// Of course this pointer might be tagged...
selectee = UNTAG_CLOSURE(val);
goto selector_loop;
}
case AP:
case AP_STACK:
case THUNK:
case THUNK_1_0:
case THUNK_0_1:
case THUNK_2_0:
case THUNK_1_1:
case THUNK_0_2:
case THUNK_STATIC:
// not evaluated yet
goto bale_out;
default:
barf("eval_thunk_selector: strange selectee %d",
(int)(info->type));
}
bale_out:
// We didn't manage to evaluate this thunk; restore the old info
// pointer. But don't forget: we still need to evacuate the thunk itself.
SET_INFO((StgClosure *)p, (const StgInfoTable *)info_ptr);
// THREADED_RTS: we just unlocked the thunk, so another thread
// might get in and update it. copy() will lock it again and
// check whether it was updated in the meantime.
*q = (StgClosure *)p;
if (evac) {
copy(q,(const StgInfoTable *)info_ptr,(StgClosure *)p,THUNK_SELECTOR_sizeW(),bd->dest_no);
}
unchain_thunk_selectors(prev_thunk_selector, *q);
return;
}
REGPARM1 GNUC_ATTR_HOT void
evacuate(StgClosure **p)
{
bdescr *bd = NULL;
nat gen_no;
StgClosure *q;
const StgInfoTable *info;
StgWord tag;
q = *p;
loop:
/* The tag and the pointer are split, to be merged after evacing */
tag = GET_CLOSURE_TAG(q);
q = UNTAG_CLOSURE(q);
ASSERTM(LOOKS_LIKE_CLOSURE_PTR(q), "invalid closure, info=%p", q->header.info);
if (!HEAP_ALLOCED_GC(q)) {
if (!major_gc) return;
info = get_itbl(q);
switch (info->type) {
case THUNK_STATIC:
if (info->srt_bitmap != 0) {
evacuate_static_object(THUNK_STATIC_LINK((StgClosure *)q), q);
}
return;
case FUN_STATIC:
if (info->srt_bitmap != 0) {
evacuate_static_object(FUN_STATIC_LINK((StgClosure *)q), q);
}
return;
case IND_STATIC:
/* If q->saved_info != NULL, then it's a revertible CAF - it'll be
* on the CAF list, so don't do anything with it here (we'll
* scavenge it later).
*/
evacuate_static_object(IND_STATIC_LINK((StgClosure *)q), q);
return;
case CONSTR_STATIC:
evacuate_static_object(STATIC_LINK(info,(StgClosure *)q), q);
return;
case CONSTR_NOCAF_STATIC:
/* no need to put these on the static linked list, they don't need
* to be scavenged.
*/
return;
default:
barf("evacuate(static): strange closure type %d", (int)(info->type));
}
}
bd = Bdescr((P_)q);
if ((bd->flags & (BF_LARGE | BF_MARKED | BF_EVACUATED)) != 0) {
// pointer into to-space: just return it. It might be a pointer
// into a generation that we aren't collecting (> N), or it
// might just be a pointer into to-space. The latter doesn't
// happen often, but allowing it makes certain things a bit
// easier; e.g. scavenging an object is idempotent, so it's OK to
// have an object on the mutable list multiple times.
if (bd->flags & BF_EVACUATED) {
// We aren't copying this object, so we have to check
// whether it is already in the target generation. (this is
// the write barrier).
if (bd->gen_no < gct->evac_gen_no) {
gct->failed_to_evac = rtsTrue;
TICK_GC_FAILED_PROMOTION();
}
return;
}
/* evacuate large objects by re-linking them onto a different list.
*/
if (bd->flags & BF_LARGE) {
evacuate_large((P_)q);
return;
}
/* If the object is in a gen that we're compacting, then we
* need to use an alternative evacuate procedure.
*/
if (!is_marked((P_)q,bd)) {
mark((P_)q,bd);
push_mark_stack((P_)q);
}
return;
}
gen_no = bd->dest_no;
info = q->header.info;
if (IS_FORWARDING_PTR(info))
{
/* Already evacuated, just return the forwarding address.
* 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 gct->failed_to_evac flag to indicate that we couldn't
* manage to promote the object to the desired generation.
*/
/*
* Optimisation: the check is fairly expensive, but we can often
* shortcut it if either the required generation is 0, or the
* current object (the EVACUATED) is in a high enough generation.
* We know that an EVACUATED always points to an object in the
* same or an older generation. gen is the lowest generation that the
* current object would be evacuated to, so we only do the full
* check if gen is too low.
*/
StgClosure *e = (StgClosure*)UN_FORWARDING_PTR(info);
*p = TAG_CLOSURE(tag,e);
if (gen_no < gct->evac_gen_no) { // optimisation
if (Bdescr((P_)e)->gen_no < gct->evac_gen_no) {
gct->failed_to_evac = rtsTrue;
TICK_GC_FAILED_PROMOTION();
}
}
return;
}
switch (INFO_PTR_TO_STRUCT(info)->type) {
case WHITEHOLE:
goto loop;
// For ints and chars of low value, save space by replacing references to
// these with closures with references to common, shared ones in the RTS.
//
// * Except when compiling into Windows DLLs which don't support cross-package
// data references very well.
//
case CONSTR_0_1:
{
#if defined(COMPILING_WINDOWS_DLL)
copy_tag_nolock(p,info,q,sizeofW(StgHeader)+1,gen_no,tag);
#else
StgWord w = (StgWord)q->payload[0];
if (info == Czh_con_info &&
// unsigned, so always true: (StgChar)w >= MIN_CHARLIKE &&
(StgChar)w <= MAX_CHARLIKE) {
*p = TAG_CLOSURE(tag,
(StgClosure *)CHARLIKE_CLOSURE((StgChar)w)
);
}
else if (info == Izh_con_info &&
(StgInt)w >= MIN_INTLIKE && (StgInt)w <= MAX_INTLIKE) {
*p = TAG_CLOSURE(tag,
(StgClosure *)INTLIKE_CLOSURE((StgInt)w)
);
}
else {
copy_tag_nolock(p,info,q,sizeofW(StgHeader)+1,gen_no,tag);
}
#endif
return;
}
case FUN_0_1:
case FUN_1_0:
case CONSTR_1_0:
copy_tag_nolock(p,info,q,sizeofW(StgHeader)+1,gen_no,tag);
return;
case THUNK_1_0:
case THUNK_0_1:
copy(p,info,q,sizeofW(StgThunk)+1,gen_no);
return;
case THUNK_1_1:
case THUNK_2_0:
case THUNK_0_2:
#ifdef NO_PROMOTE_THUNKS
#error bitrotted
#endif
copy(p,info,q,sizeofW(StgThunk)+2,gen_no);
return;
case FUN_1_1:
case FUN_2_0:
case FUN_0_2:
case CONSTR_1_1:
case CONSTR_2_0:
copy_tag_nolock(p,info,q,sizeofW(StgHeader)+2,gen_no,tag);
return;
case CONSTR_0_2:
copy_tag_nolock(p,info,q,sizeofW(StgHeader)+2,gen_no,tag);
return;
case THUNK:
copy(p,info,q,thunk_sizeW_fromITBL(INFO_PTR_TO_STRUCT(info)),gen_no);
return;
case FUN:
case CONSTR:
copy_tag_nolock(p,info,q,sizeW_fromITBL(INFO_PTR_TO_STRUCT(info)),gen_no,tag);
return;
case BLACKHOLE:
{
StgClosure *r;
const StgInfoTable *i;
r = ((StgInd*)q)->indirectee;
if (GET_CLOSURE_TAG(r) == 0) {
i = r->header.info;
if (IS_FORWARDING_PTR(i)) {
r = (StgClosure *)UN_FORWARDING_PTR(i);
i = r->header.info;
}
if (i == &stg_TSO_info
|| i == &stg_WHITEHOLE_info
|| i == &stg_BLOCKING_QUEUE_CLEAN_info
|| i == &stg_BLOCKING_QUEUE_DIRTY_info) {
copy(p,info,q,sizeofW(StgInd),gen_no);
return;
}
ASSERT(i != &stg_IND_info);
}
q = r;
*p = r;
goto loop;
}
case MUT_VAR_CLEAN:
case MUT_VAR_DIRTY:
case MVAR_CLEAN:
case MVAR_DIRTY:
case TVAR:
case BLOCKING_QUEUE:
case WEAK:
case PRIM:
case MUT_PRIM:
copy(p,info,q,sizeW_fromITBL(INFO_PTR_TO_STRUCT(info)),gen_no);
return;
case BCO:
copy(p,info,q,bco_sizeW((StgBCO *)q),gen_no);
return;
case THUNK_SELECTOR:
eval_thunk_selector(p, (StgSelector *)q, rtsTrue);
return;
case IND:
// follow chains of indirections, don't evacuate them
q = ((StgInd*)q)->indirectee;
*p = q;
goto loop;
case RET_BCO:
case RET_SMALL:
case RET_BIG:
case UPDATE_FRAME:
case UNDERFLOW_FRAME:
case STOP_FRAME:
case CATCH_FRAME:
case CATCH_STM_FRAME:
case CATCH_RETRY_FRAME:
case ATOMICALLY_FRAME:
// shouldn't see these
barf("evacuate: stack frame at %p\n", q);
case PAP:
copy(p,info,q,pap_sizeW((StgPAP*)q),gen_no);
return;
case AP:
copy(p,info,q,ap_sizeW((StgAP*)q),gen_no);
return;
case AP_STACK:
copy(p,info,q,ap_stack_sizeW((StgAP_STACK*)q),gen_no);
return;
case ARR_WORDS:
// just copy the block
copy(p,info,q,arr_words_sizeW((StgArrBytes *)q),gen_no);
return;
case MUT_ARR_PTRS_CLEAN:
case MUT_ARR_PTRS_DIRTY:
case MUT_ARR_PTRS_FROZEN:
case MUT_ARR_PTRS_FROZEN0:
// just copy the block
copy(p,info,q,mut_arr_ptrs_sizeW((StgMutArrPtrs *)q),gen_no);
return;
case SMALL_MUT_ARR_PTRS_CLEAN:
case SMALL_MUT_ARR_PTRS_DIRTY:
case SMALL_MUT_ARR_PTRS_FROZEN:
case SMALL_MUT_ARR_PTRS_FROZEN0:
// just copy the block
copy(p,info,q,small_mut_arr_ptrs_sizeW((StgSmallMutArrPtrs *)q),gen_no);
return;
case TSO:
copy(p,info,q,sizeofW(StgTSO),gen_no);
return;
case STACK:
{
StgStack *stack = (StgStack *)q;
/* To evacuate a small STACK, we need to adjust the stack pointer
*/
{
StgStack *new_stack;
StgPtr r, s;
rtsBool mine;
mine = copyPart(p,(StgClosure *)stack, stack_sizeW(stack),
sizeofW(StgStack), gen_no);
if (mine) {
new_stack = (StgStack *)*p;
move_STACK(stack, new_stack);
for (r = stack->sp, s = new_stack->sp;
r < stack->stack + stack->stack_size;) {
*s++ = *r++;
}
}
return;
}
}
case TREC_CHUNK:
copy(p,info,q,sizeofW(StgTRecChunk),gen_no);
return;
default:
barf("evacuate: strange closure type %d", (int)(INFO_PTR_TO_STRUCT(info)->type));
}
barf("evacuate");
}
StgClosure *
isAlive(StgClosure *p)
{
const StgInfoTable *info;
bdescr *bd;
StgWord tag;
StgClosure *q;
while (1) {
/* The tag and the pointer are split, to be merged later when needed. */
tag = GET_CLOSURE_TAG(p);
q = UNTAG_CLOSURE(p);
ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
// ignore static closures
//
// ToDo: This means we never look through IND_STATIC, which means
// isRetainer needs to handle the IND_STATIC case rather than
// raising an error.
//
// ToDo: for static closures, check the static link field.
// Problem here is that we sometimes don't set the link field, eg.
// for static closures with an empty SRT or CONSTR_STATIC_NOCAFs.
//
if (!HEAP_ALLOCED_GC(q)) {
return p;
}
// ignore closures in generations that we're not collecting.
bd = Bdescr((P_)q);
// if it's a pointer into to-space, then we're done
if (bd->flags & BF_EVACUATED) {
return p;
}
// large objects use the evacuated flag
if (bd->flags & BF_LARGE) {
return NULL;
}
// check the mark bit for compacted steps
if ((bd->flags & BF_MARKED) && is_marked((P_)q,bd)) {
return p;
}
info = q->header.info;
if (IS_FORWARDING_PTR(info)) {
// alive!
return TAG_CLOSURE(tag,(StgClosure*)UN_FORWARDING_PTR(info));
}
info = INFO_PTR_TO_STRUCT(info);
switch (info->type) {
case IND:
case IND_STATIC:
case IND_PERM:
// follow indirections
p = ((StgInd *)q)->indirectee;
continue;
case BLACKHOLE:
p = ((StgInd*)q)->indirectee;
if (GET_CLOSURE_TAG(p) != 0) {
continue;
} else {
return NULL;
}
default:
// dead.
return NULL;
}
}
}
void
pruneSparkQueue (Capability *cap)
{
SparkPool *pool;
StgClosurePtr spark, tmp, *elements;
nat n, pruned_sparks; // stats only
StgWord botInd,oldBotInd,currInd; // indices in array (always < size)
const StgInfoTable *info;
n = 0;
pruned_sparks = 0;
pool = cap->sparks;
// it is possible that top > bottom, indicating an empty pool. We
// fix that here; this is only necessary because the loop below
// assumes it.
if (pool->top > pool->bottom)
pool->top = pool->bottom;
// Take this opportunity to reset top/bottom modulo the size of
// the array, to avoid overflow. This is only possible because no
// stealing is happening during GC.
pool->bottom -= pool->top & ~pool->moduloSize;
pool->top &= pool->moduloSize;
pool->topBound = pool->top;
debugTrace(DEBUG_sparks,
"markSparkQueue: current spark queue len=%ld; (hd=%ld; tl=%ld)",
sparkPoolSize(pool), pool->bottom, pool->top);
ASSERT_WSDEQUE_INVARIANTS(pool);
elements = (StgClosurePtr *)pool->elements;
/* We have exclusive access to the structure here, so we can reset
bottom and top counters, and prune invalid sparks. Contents are
copied in-place if they are valuable, otherwise discarded. The
routine uses "real" indices t and b, starts by computing them
as the modulus size of top and bottom,
Copying:
At the beginning, the pool structure can look like this:
( bottom % size >= top % size , no wrap-around)
t b
___________***********_________________
or like this ( bottom % size < top % size, wrap-around )
b t
***********__________******************
As we need to remove useless sparks anyway, we make one pass
between t and b, moving valuable content to b and subsequent
cells (wrapping around when the size is reached).
b t
***********OOO_______XX_X__X?**********
^____move?____/
After this movement, botInd becomes the new bottom, and old
bottom becomes the new top index, both as indices in the array
size range.
*/
// starting here
currInd = (pool->top) & (pool->moduloSize); // mod
// copies of evacuated closures go to space from botInd on
// we keep oldBotInd to know when to stop
oldBotInd = botInd = (pool->bottom) & (pool->moduloSize); // mod
// on entry to loop, we are within the bounds
ASSERT( currInd < pool->size && botInd < pool->size );
while (currInd != oldBotInd ) {
/* must use != here, wrap-around at size
subtle: loop not entered if queue empty
*/
/* check element at currInd. if valuable, evacuate and move to
botInd, otherwise move on */
spark = elements[currInd];
// We have to be careful here: in the parallel GC, another
// thread might evacuate this closure while we're looking at it,
// so grab the info pointer just once.
if (GET_CLOSURE_TAG(spark) != 0) {
// Tagged pointer is a value, so the spark has fizzled. It
// probably never happens that we get a tagged pointer in
// the spark pool, because we would have pruned the spark
// during the previous GC cycle if it turned out to be
// evaluated, but it doesn't hurt to have this check for
// robustness.
pruned_sparks++;
cap->sparks_fizzled++;
} else {
info = spark->header.info;
if (IS_FORWARDING_PTR(info)) {
tmp = (StgClosure*)UN_FORWARDING_PTR(info);
/* if valuable work: shift inside the pool */
if (closure_SHOULD_SPARK(tmp)) {
elements[botInd] = tmp; // keep entry (new address)
botInd++;
n++;
} else {
pruned_sparks++; // discard spark
cap->sparks_fizzled++;
}
} else if (HEAP_ALLOCED(spark)) {
if ((Bdescr((P_)spark)->flags & BF_EVACUATED)) {
if (closure_SHOULD_SPARK(spark)) {
elements[botInd] = spark; // keep entry (new address)
botInd++;
n++;
} else {
pruned_sparks++; // discard spark
cap->sparks_fizzled++;
}
} else {
pruned_sparks++; // discard spark
cap->sparks_gcd++;
}
} else {
if (INFO_PTR_TO_STRUCT(info)->type == THUNK_STATIC) {
if (*THUNK_STATIC_LINK(spark) != NULL) {
elements[botInd] = spark; // keep entry (new address)
botInd++;
n++;
} else {
pruned_sparks++; // discard spark
cap->sparks_gcd++;
}
} else {
pruned_sparks++; // discard spark
cap->sparks_fizzled++;
}
}
}
currInd++;
// in the loop, we may reach the bounds, and instantly wrap around
ASSERT( currInd <= pool->size && botInd <= pool->size );
if ( currInd == pool->size ) { currInd = 0; }
if ( botInd == pool->size ) { botInd = 0; }
} // while-loop over spark pool elements
ASSERT(currInd == oldBotInd);
pool->top = oldBotInd; // where we started writing
pool->topBound = pool->top;
pool->bottom = (oldBotInd <= botInd) ? botInd : (botInd + pool->size);
// first free place we did not use (corrected by wraparound)
debugTrace(DEBUG_sparks, "pruned %d sparks", pruned_sparks);
debugTrace(DEBUG_sparks,
"new spark queue len=%ld; (hd=%ld; tl=%ld)",
sparkPoolSize(pool), pool->bottom, pool->top);
ASSERT_WSDEQUE_INVARIANTS(pool);
}
static nat
update_bkwd_compact( step *stp )
{
StgPtr p, free;
#if 0
StgWord m;
#endif
bdescr *bd, *free_bd;
StgInfoTable *info;
nat size, free_blocks;
StgWord iptr;
bd = free_bd = stp->old_blocks;
free = free_bd->start;
free_blocks = 1;
// cycle through all the blocks in the step
for (; bd != NULL; bd = bd->link) {
p = bd->start;
while (p < bd->free ) {
while ( p < bd->free && !is_marked(p,bd) ) {
p++;
}
if (p >= bd->free) {
break;
}
#if 0
next:
m = * ((StgPtr)bd->u.bitmap + ((p - bd->start) / (BITS_IN(StgWord))));
m >>= ((p - bd->start) & (BITS_IN(StgWord) - 1));
while ( p < bd->free ) {
if ((m & 1) == 0) {
m >>= 1;
p++;
if (((StgWord)p & (sizeof(W_) * BITS_IN(StgWord))) == 0) {
goto next;
} else {
continue;
}
}
#endif
if (!is_marked(p+1,bd)) {
// don't forget to update the free ptr in the block desc.
free_bd->free = free;
free_bd = free_bd->link;
free = free_bd->start;
free_blocks++;
}
iptr = get_threaded_info(p);
unthread(p, (StgWord)free + GET_CLOSURE_TAG((StgClosure *)iptr));
ASSERT(LOOKS_LIKE_INFO_PTR(((StgClosure *)p)->header.info));
info = get_itbl((StgClosure *)p);
size = closure_sizeW_((StgClosure *)p,info);
if (free != p) {
move(free,p,size);
}
// relocate TSOs
if (info->type == TSO) {
move_TSO((StgTSO *)p, (StgTSO *)free);
}
free += size;
p += size;
#if 0
goto next;
#endif
}
}
// free the remaining blocks and count what's left.
free_bd->free = free;
if (free_bd->link != NULL) {
freeChain(free_bd->link);
free_bd->link = NULL;
}
return free_blocks;
}
static void
update_fwd_compact( bdescr *blocks )
{
StgPtr p, q, free;
#if 0
StgWord m;
#endif
bdescr *bd, *free_bd;
StgInfoTable *info;
nat size;
StgWord iptr;
bd = blocks;
free_bd = blocks;
free = free_bd->start;
// cycle through all the blocks in the step
for (; bd != NULL; bd = bd->link) {
p = bd->start;
while (p < bd->free ) {
while ( p < bd->free && !is_marked(p,bd) ) {
p++;
}
if (p >= bd->free) {
break;
}
#if 0
next:
m = * ((StgPtr)bd->u.bitmap + ((p - bd->start) / (BITS_IN(StgWord))));
m >>= ((p - bd->start) & (BITS_IN(StgWord) - 1));
while ( p < bd->free ) {
if ((m & 1) == 0) {
m >>= 1;
p++;
if (((StgWord)p & (sizeof(W_) * BITS_IN(StgWord))) == 0) {
goto next;
} else {
continue;
}
}
#endif
// Problem: we need to know the destination for this cell
// in order to unthread its info pointer. But we can't
// know the destination without the size, because we may
// spill into the next block. So we have to run down the
// threaded list and get the info ptr first.
//
// ToDo: one possible avenue of attack is to use the fact
// that if (p&BLOCK_MASK) >= (free&BLOCK_MASK), then we
// definitely have enough room. Also see bug #1147.
iptr = get_threaded_info(p);
info = INFO_PTR_TO_STRUCT(UNTAG_CLOSURE((StgClosure *)iptr));
q = p;
p = thread_obj(info, p);
size = p - q;
if (free + size > free_bd->start + BLOCK_SIZE_W) {
// unset the next bit in the bitmap to indicate that
// this object needs to be pushed into the next
// block. This saves us having to run down the
// threaded info pointer list twice during the next pass.
unmark(q+1,bd);
free_bd = free_bd->link;
free = free_bd->start;
} else {
ASSERT(is_marked(q+1,bd));
}
unthread(q,(StgWord)free + GET_CLOSURE_TAG((StgClosure *)iptr));
free += size;
#if 0
goto next;
#endif
}
}
}
static void
scavenge_stack(StgPtr p, StgPtr stack_end)
{
const StgRetInfoTable* info;
StgWord bitmap;
nat size;
/*
* Each time around this loop, we are looking at a chunk of stack
* that starts with an activation record.
*/
while (p < stack_end) {
info = get_ret_itbl((StgClosure *)p);
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
// check whether the value after evacuation is a BLACKHOLE,
// and if not, we change the update frame to an stg_enter
// frame that simply returns the value. Hence, blackholing is
// compulsory (otherwise we would have to check for thunks
// too).
//
// Note [upd-black-hole]
// One slight hiccup is that the THUNK_SELECTOR machinery can
// overwrite the updatee with an IND. In parallel GC, this
// could even be happening concurrently, so we can't check for
// the IND. Fortunately if we assume that blackholing is
// happening (either lazy or eager), then we can be sure that
// the updatee is never a THUNK_SELECTOR and we're ok.
// NB. this is a new invariant: blackholing is not optional.
{
StgUpdateFrame *frame = (StgUpdateFrame *)p;
StgClosure *v;
evacuate(&frame->updatee);
v = frame->updatee;
if (GET_CLOSURE_TAG(v) != 0 ||
(get_itbl(v)->type != BLACKHOLE)) {
// blackholing is compulsory, see above.
frame->header.info = (const StgInfoTable*)&stg_enter_checkbh_info;
}
ASSERT(v->header.info != &stg_TSO_info);
p += sizeofW(StgUpdateFrame);
continue;
}
// small bitmap (< 32 entries, or 64 on a 64-bit machine)
case CATCH_STM_FRAME:
case CATCH_RETRY_FRAME:
case ATOMICALLY_FRAME:
case UNDERFLOW_FRAME:
case STOP_FRAME:
case CATCH_FRAME:
case RET_SMALL:
bitmap = BITMAP_BITS(info->i.layout.bitmap);
size = BITMAP_SIZE(info->i.layout.bitmap);
// NOTE: the payload starts immediately after the info-ptr, we
// don't have an StgHeader in the same sense as a heap closure.
p++;
p = scavenge_small_bitmap(p, size, bitmap);
follow_srt:
if (major_gc)
scavenge_srt((StgClosure **)GET_SRT(info), info->i.srt_bitmap);
continue;
case RET_BCO: {
StgBCO *bco;
nat size;
p++;
evacuate((StgClosure **)p);
bco = (StgBCO *)*p;
p++;
size = BCO_BITMAP_SIZE(bco);
scavenge_large_bitmap(p, BCO_BITMAP(bco), size);
p += size;
continue;
}
// large bitmap (> 32 entries, or > 64 on a 64-bit machine)
case RET_BIG:
{
nat size;
size = GET_LARGE_BITMAP(&info->i)->size;
p++;
scavenge_large_bitmap(p, GET_LARGE_BITMAP(&info->i), size);
p += size;
// and don't forget to follow the SRT
goto follow_srt;
}
// Dynamic bitmap: the mask is stored on the stack, and
// there are a number of non-pointers followed by a number
// of pointers above the bitmapped area. (see StgMacros.h,
// HEAP_CHK_GEN).
case RET_DYN:
{
StgWord dyn;
dyn = ((StgRetDyn *)p)->liveness;
// traverse the bitmap first
bitmap = RET_DYN_LIVENESS(dyn);
p = (P_)&((StgRetDyn *)p)->payload[0];
size = RET_DYN_BITMAP_SIZE;
p = scavenge_small_bitmap(p, size, bitmap);
// skip over the non-ptr words
p += RET_DYN_NONPTRS(dyn) + RET_DYN_NONPTR_REGS_SIZE;
// follow the ptr words
for (size = RET_DYN_PTRS(dyn); size > 0; size--) {
evacuate((StgClosure **)p);
p++;
}
continue;
}
case RET_FUN:
{
StgRetFun *ret_fun = (StgRetFun *)p;
StgFunInfoTable *fun_info;
evacuate(&ret_fun->fun);
fun_info = get_fun_itbl(UNTAG_CLOSURE(ret_fun->fun));
p = scavenge_arg_block(fun_info, ret_fun->payload);
goto follow_srt;
}
default:
barf("scavenge_stack: weird activation record found on stack: %d", (int)(info->i.type));
}
}
}
