/* -----------------------------------------------------------------------------
- * $Id: GC.c,v 1.117 2001/08/08 13:45:02 simonmar Exp $
+ * $Id: GC.c,v 1.146 2002/12/11 15:36:42 simonmar Exp $
*
- * (c) The GHC Team 1998-1999
+ * (c) The GHC Team 1998-2002
*
* Generational garbage collector
*
* ---------------------------------------------------------------------------*/
+#include "PosixSource.h"
#include "Rts.h"
#include "RtsFlags.h"
#include "RtsUtils.h"
+#include "Apply.h"
#include "Storage.h"
#include "StoragePriv.h"
#include "Stats.h"
#include "Prelude.h"
#include "ParTicky.h" // ToDo: move into Rts.h
#include "GCCompact.h"
+#include "Signals.h"
#if defined(GRAN) || defined(PAR)
# include "GranSimRts.h"
# include "ParallelRts.h"
#include "FrontPanel.h"
#endif
+#include "RetainerProfile.h"
+#include "LdvProfile.h"
+
+#include <string.h>
+
/* STATIC OBJECT LIST.
*
* During GC:
* We build up a static object list while collecting generations 0..N,
* which is then appended to the static object list of generation N+1.
*/
-StgClosure* static_objects; // live static objects
-StgClosure* scavenged_static_objects; // static objects scavenged so far
+static StgClosure* static_objects; // live static objects
+StgClosure* scavenged_static_objects; // static objects scavenged so far
/* N is the oldest generation being collected, where the generations
* are numbered starting at 0. A major GC (indicated by the major_gc
/* Weak pointers
*/
StgWeak *old_weak_ptr_list; // also pending finaliser list
-static rtsBool weak_done; // all done for this pass
+
+/* Which stage of processing various kinds of weak pointer are we at?
+ * (see traverse_weak_ptr_list() below for discussion).
+ */
+typedef enum { WeakPtrs, WeakThreads, WeakDone } WeakStage;
+static WeakStage weak_stage;
/* List of all threads during GC
*/
static StgTSO *old_all_threads;
-static StgTSO *resurrected_threads;
+StgTSO *resurrected_threads;
/* Flag indicating failure to evacuate an object to the desired
* generation.
/* Old to-space (used for two-space collector only)
*/
-bdescr *old_to_blocks;
+static bdescr *old_to_blocks;
/* Data used for allocation area sizing.
*/
-lnat new_blocks; // blocks allocated during this GC
-lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
+static lnat new_blocks; // blocks allocated during this GC
+static lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
/* Used to avoid long recursion due to selector thunks
*/
-lnat thunk_selector_depth = 0;
-#define MAX_THUNK_SELECTOR_DEPTH 256
+static lnat thunk_selector_depth = 0;
+#define MAX_THUNK_SELECTOR_DEPTH 8
/* -----------------------------------------------------------------------------
Static function declarations
-------------------------------------------------------------------------- */
+static bdescr * gc_alloc_block ( step *stp );
static void mark_root ( StgClosure **root );
static StgClosure * evacuate ( StgClosure *q );
static void zero_static_object_list ( StgClosure* first_static );
static rtsBool traverse_weak_ptr_list ( void );
static void mark_weak_ptr_list ( StgWeak **list );
-static void scavenge ( step * );
-static void scavenge_mark_stack ( void );
-static void scavenge_stack ( StgPtr p, StgPtr stack_end );
-static rtsBool scavenge_one ( StgPtr p );
-static void scavenge_large ( step * );
-static void scavenge_static ( void );
-static void scavenge_mutable_list ( generation *g );
-static void scavenge_mut_once_list ( generation *g );
-static void scavengeCAFs ( void );
+static StgClosure * eval_thunk_selector ( nat field, StgSelector * p );
+
+
+static void scavenge ( step * );
+static void scavenge_mark_stack ( void );
+static void scavenge_stack ( StgPtr p, StgPtr stack_end );
+static rtsBool scavenge_one ( StgPtr p );
+static void scavenge_large ( step * );
+static void scavenge_static ( void );
+static void scavenge_mutable_list ( generation *g );
+static void scavenge_mut_once_list ( generation *g );
+
+static void scavenge_large_bitmap ( StgPtr p,
+ StgLargeBitmap *large_bitmap,
+ nat size );
#if 0 && defined(DEBUG)
static void gcCAFs ( void );
}
/* -----------------------------------------------------------------------------
+ Allocate a new to-space block in the given step.
+ -------------------------------------------------------------------------- */
+
+static bdescr *
+gc_alloc_block(step *stp)
+{
+ bdescr *bd = allocBlock();
+ bd->gen_no = stp->gen_no;
+ bd->step = stp;
+ bd->link = NULL;
+
+ // blocks in to-space in generations up to and including N
+ // get the BF_EVACUATED flag.
+ if (stp->gen_no <= N) {
+ bd->flags = BF_EVACUATED;
+ } else {
+ bd->flags = 0;
+ }
+
+ // Start a new to-space block, chain it on after the previous one.
+ if (stp->hp_bd == NULL) {
+ stp->hp_bd = bd;
+ } else {
+ stp->hp_bd->free = stp->hp;
+ stp->hp_bd->link = bd;
+ stp->hp_bd = bd;
+ }
+
+ stp->hp = bd->start;
+ stp->hpLim = stp->hp + BLOCK_SIZE_W;
+
+ stp->n_to_blocks++;
+ new_blocks++;
+
+ return bd;
+}
+
+/* -----------------------------------------------------------------------------
GarbageCollect
- For garbage collecting generation N (and all younger generations):
+ Rough outline of the algorithm: for garbage collecting generation N
+ (and all younger generations):
- follow all pointers in the root set. the root set includes all
- mutable objects in all steps in all generations.
+ mutable objects in all generations (mutable_list and mut_once_list).
- for each pointer, evacuate the object it points to into either
- + to-space in the next higher step in that generation, if one exists,
- + if the object's generation == N, then evacuate it to the next
- generation if one exists, or else to-space in the current
- generation.
- + if the object's generation < N, then evacuate it to to-space
- in the next generation.
+
+ + to-space of the step given by step->to, which is the next
+ highest step in this generation or the first step in the next
+ generation if this is the last step.
+
+ + to-space of generations[evac_gen]->steps[0], if evac_gen != 0.
+ When we evacuate an object we attempt to evacuate
+ everything it points to into the same generation - this is
+ achieved by setting evac_gen to the desired generation. If
+ we can't do this, then an entry in the mut_once list has to
+ be made for the cross-generation pointer.
+
+ + if the object is already in a generation > N, then leave
+ it alone.
- repeatedly scavenge to-space from each step in each generation
being collected until no more objects can be evacuated.
- free from-space in each step, and set from-space = to-space.
+ Locks held: sched_mutex
+
-------------------------------------------------------------------------- */
void
Now, Now));
#endif
+#ifndef mingw32_TARGET_OS
+ // block signals
+ blockUserSignals();
+#endif
+
// tell the stats department that we've started a GC
stat_startGC();
*/
new_blocks = 0;
- /* Initialise to-space in all the generations/steps that we're
- * collecting.
- */
+ // Initialise to-space in all the generations/steps that we're
+ // collecting.
+ //
for (g = 0; g <= N; g++) {
generations[g].mut_once_list = END_MUT_LIST;
generations[g].mut_list = END_MUT_LIST;
continue;
}
- /* Get a free block for to-space. Extra blocks will be chained on
- * as necessary.
- */
- bd = allocBlock();
stp = &generations[g].steps[s];
ASSERT(stp->gen_no == g);
- ASSERT(stp->hp ? Bdescr(stp->hp)->step == stp : rtsTrue);
- bd->gen_no = g;
- bd->step = stp;
- bd->link = NULL;
- bd->flags = BF_EVACUATED; // it's a to-space block
- stp->hp = bd->start;
- stp->hpLim = stp->hp + BLOCK_SIZE_W;
- stp->hp_bd = bd;
+
+ // start a new to-space for this step.
+ stp->hp = NULL;
+ stp->hp_bd = NULL;
+ stp->to_blocks = NULL;
+
+ // allocate the first to-space block; extra blocks will be
+ // chained on as necessary.
+ bd = gc_alloc_block(stp);
stp->to_blocks = bd;
- stp->n_to_blocks = 1;
stp->scan = bd->start;
stp->scan_bd = bd;
+
+ // initialise the large object queues.
stp->new_large_objects = NULL;
stp->scavenged_large_objects = NULL;
stp->n_scavenged_large_blocks = 0;
- new_blocks++;
+
// mark the large objects as not evacuated yet
for (bd = stp->large_objects; bd; bd = bd->link) {
bd->flags = BF_LARGE;
}
/* make sure the older generations have at least one block to
- * allocate into (this makes things easier for copy(), see below.
+ * allocate into (this makes things easier for copy(), see below).
*/
for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
for (s = 0; s < generations[g].n_steps; s++) {
stp = &generations[g].steps[s];
if (stp->hp_bd == NULL) {
ASSERT(stp->blocks == NULL);
- bd = allocBlock();
- bd->gen_no = g;
- bd->step = stp;
- bd->link = NULL;
- bd->flags = 0; // *not* a to-space block or a large object
- stp->hp = bd->start;
- stp->hpLim = stp->hp + BLOCK_SIZE_W;
- stp->hp_bd = bd;
+ bd = gc_alloc_block(stp);
stp->blocks = bd;
stp->n_blocks = 1;
- new_blocks++;
}
/* Set the scan pointer for older generations: remember we
* still have to scavenge objects that have been promoted. */
}
}
- scavengeCAFs();
+ /* follow roots from the CAF list (used by GHCi)
+ */
+ evac_gen = 0;
+ markCAFs(mark_root);
/* follow all the roots that the application knows about.
*/
mark_weak_ptr_list(&weak_ptr_list);
old_weak_ptr_list = weak_ptr_list;
weak_ptr_list = NULL;
- weak_done = rtsFalse;
+ weak_stage = WeakPtrs;
/* The all_threads list is like the weak_ptr_list.
* See traverse_weak_ptr_list() for the details.
if (flag) { goto loop; }
- // must be last...
+ // must be last... invariant is that everything is fully
+ // scavenged at this point.
if (traverse_weak_ptr_list()) { // returns rtsTrue if evaced something
goto loop;
}
}
+ /* Update the pointers from the "main thread" list - these are
+ * treated as weak pointers because we want to allow a main thread
+ * to get a BlockedOnDeadMVar exception in the same way as any other
+ * thread. Note that the threads should all have been retained by
+ * GC by virtue of being on the all_threads list, we're just
+ * updating pointers here.
+ */
+ {
+ StgMainThread *m;
+ StgTSO *tso;
+ for (m = main_threads; m != NULL; m = m->link) {
+ tso = (StgTSO *) isAlive((StgClosure *)m->tso);
+ if (tso == NULL) {
+ barf("main thread has been GC'd");
+ }
+ m->tso = tso;
+ }
+ }
+
#if defined(PAR)
// Reconstruct the Global Address tables used in GUM
rebuildGAtables(major_gc);
for (s = 0; s < generations[g].n_steps; s++) {
stp = &generations[g].steps[s];
if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
+ ASSERT(Bdescr(stp->hp) == stp->hp_bd);
stp->hp_bd->free = stp->hp;
- stp->hp_bd->link = NULL;
}
}
}
+#ifdef PROFILING
+ // We call processHeapClosureForDead() on every closure destroyed during
+ // the current garbage collection, so we invoke LdvCensusForDead().
+ if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
+ || RtsFlags.ProfFlags.bioSelector != NULL)
+ LdvCensusForDead(N);
+#endif
+
// NO MORE EVACUATION AFTER THIS POINT!
// Finally: compaction of the oldest generation.
if (major_gc && oldest_gen->steps[0].is_compacted) {
// add the new blocks we promoted during this GC
stp->n_blocks += stp->n_to_blocks;
+ stp->n_to_blocks = 0;
stp->n_large_blocks += stp->n_scavenged_large_blocks;
}
}
RtsFlags.GcFlags.minOldGenSize);
// minimum size for generation zero
- min_alloc = (RtsFlags.GcFlags.pcFreeHeap * max) / 200;
+ min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
+ RtsFlags.GcFlags.minAllocAreaSize);
+
+ // Auto-enable compaction when the residency reaches a
+ // certain percentage of the maximum heap size (default: 30%).
+ if (RtsFlags.GcFlags.generations > 1 &&
+ (RtsFlags.GcFlags.compact ||
+ (max > 0 &&
+ oldest_gen->steps[0].n_blocks >
+ (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
+ oldest_gen->steps[0].is_compacted = 1;
+// fprintf(stderr,"compaction: on\n", live);
+ } else {
+ oldest_gen->steps[0].is_compacted = 0;
+// fprintf(stderr,"compaction: off\n", live);
+ }
// if we're going to go over the maximum heap size, reduce the
// size of the generations accordingly. The calculation is
// different if compaction is turned on, because we don't need
// to double the space required to collect the old generation.
if (max != 0) {
- if (RtsFlags.GcFlags.compact) {
+
+ // this test is necessary to ensure that the calculations
+ // below don't have any negative results - we're working
+ // with unsigned values here.
+ if (max < min_alloc) {
+ heapOverflow();
+ }
+
+ if (oldest_gen->steps[0].is_compacted) {
if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
size = (max - min_alloc) / ((gens - 1) * 2 - 1);
}
for (g = 0; g < gens; g++) {
generations[g].max_blocks = size;
}
-
- // Auto-enable compaction when the residency reaches a
- // certain percentage of the maximum heap size (default: 30%).
- if (RtsFlags.GcFlags.compact &&
- max > 0 &&
- oldest_gen->steps[0].n_blocks >
- (RtsFlags.GcFlags.compactThreshold * max) / 100) {
- oldest_gen->steps[0].is_compacted = 1;
-// fprintf(stderr,"compaction: on\n", live);
- } else {
- oldest_gen->steps[0].is_compacted = 0;
-// fprintf(stderr,"compaction: off\n", live);
- }
}
// Guess the amount of live data for stats.
*/
blocks = g0s0->n_to_blocks;
- if ( blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
- RtsFlags.GcFlags.maxHeapSize ) {
+ if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
+ blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
+ RtsFlags.GcFlags.maxHeapSize ) {
long adjusted_blocks; // signed on purpose
int pc_free;
}
resizeNursery((nat)blocks);
+
+ } else {
+ // we might have added extra large blocks to the nursery, so
+ // resize back to minAllocAreaSize again.
+ resizeNursery(RtsFlags.GcFlags.minAllocAreaSize);
}
}
if (major_gc) { gcCAFs(); }
#endif
+#ifdef PROFILING
+ // resetStaticObjectForRetainerProfiling() must be called before
+ // zeroing below.
+ resetStaticObjectForRetainerProfiling();
+#endif
+
// zero the scavenged static object list
if (major_gc) {
zero_static_object_list(scavenged_static_objects);
// Reset the nursery
resetNurseries();
+ RELEASE_LOCK(&sched_mutex);
+
// start any pending finalizers
scheduleFinalizers(old_weak_ptr_list);
// send exceptions to any threads which were about to die
resurrectThreads(resurrected_threads);
+
+ ACQUIRE_LOCK(&sched_mutex);
// Update the stable pointer hash table.
updateStablePtrTable(major_gc);
// restore enclosing cost centre
#ifdef PROFILING
- heapCensus();
CCCS = prev_CCS;
#endif
// ok, GC over: tell the stats department what happened.
stat_endGC(allocated, collected, live, copied, N);
+#ifndef mingw32_TARGET_OS
+ // unblock signals again
+ unblockUserSignals();
+#endif
+
//PAR_TICKY_TP();
}
older generations than the one we're collecting. This could
probably be optimised by keeping per-generation lists of weak
pointers, but for a few weak pointers this scheme will work.
+
+ There are three distinct stages to processing weak pointers:
+
+ - weak_stage == WeakPtrs
+
+ We process all the weak pointers whos keys are alive (evacuate
+ their values and finalizers), and repeat until we can find no new
+ live keys. If no live keys are found in this pass, then we
+ evacuate the finalizers of all the dead weak pointers in order to
+ run them.
+
+ - weak_stage == WeakThreads
+
+ Now, we discover which *threads* are still alive. Pointers to
+ threads from the all_threads and main thread lists are the
+ weakest of all: a pointers from the finalizer of a dead weak
+ pointer can keep a thread alive. Any threads found to be unreachable
+ are evacuated and placed on the resurrected_threads list so we
+ can send them a signal later.
+
+ - weak_stage == WeakDone
+
+ No more evacuation is done.
+
-------------------------------------------------------------------------- */
static rtsBool
StgClosure *new;
rtsBool flag = rtsFalse;
- if (weak_done) { return rtsFalse; }
-
- /* doesn't matter where we evacuate values/finalizers to, since
- * these pointers are treated as roots (iff the keys are alive).
- */
- evac_gen = 0;
-
- last_w = &old_weak_ptr_list;
- for (w = old_weak_ptr_list; w != NULL; w = next_w) {
+ switch (weak_stage) {
- /* There might be a DEAD_WEAK on the list if finalizeWeak# was
- * called on a live weak pointer object. Just remove it.
- */
- if (w->header.info == &stg_DEAD_WEAK_info) {
- next_w = ((StgDeadWeak *)w)->link;
- *last_w = next_w;
- continue;
- }
-
- ASSERT(get_itbl(w)->type == WEAK);
-
- /* Now, check whether the key is reachable.
- */
- new = isAlive(w->key);
- if (new != NULL) {
- w->key = new;
- // evacuate the value and finalizer
- w->value = evacuate(w->value);
- w->finalizer = evacuate(w->finalizer);
- // remove this weak ptr from the old_weak_ptr list
- *last_w = w->link;
- // and put it on the new weak ptr list
- next_w = w->link;
- w->link = weak_ptr_list;
- weak_ptr_list = w;
- flag = rtsTrue;
- IF_DEBUG(weak, belch("Weak pointer still alive at %p -> %p", w, w->key));
- continue;
- }
- else {
- last_w = &(w->link);
- next_w = w->link;
- continue;
- }
- }
+ case WeakDone:
+ return rtsFalse;
- /* Now deal with the all_threads list, which behaves somewhat like
- * the weak ptr list. If we discover any threads that are about to
- * become garbage, we wake them up and administer an exception.
- */
- {
- StgTSO *t, *tmp, *next, **prev;
+ case WeakPtrs:
+ /* doesn't matter where we evacuate values/finalizers to, since
+ * these pointers are treated as roots (iff the keys are alive).
+ */
+ evac_gen = 0;
+
+ last_w = &old_weak_ptr_list;
+ for (w = old_weak_ptr_list; w != NULL; w = next_w) {
+
+ /* There might be a DEAD_WEAK on the list if finalizeWeak# was
+ * called on a live weak pointer object. Just remove it.
+ */
+ if (w->header.info == &stg_DEAD_WEAK_info) {
+ next_w = ((StgDeadWeak *)w)->link;
+ *last_w = next_w;
+ continue;
+ }
+
+ switch (get_itbl(w)->type) {
- prev = &old_all_threads;
- for (t = old_all_threads; t != END_TSO_QUEUE; t = next) {
+ case EVACUATED:
+ next_w = (StgWeak *)((StgEvacuated *)w)->evacuee;
+ *last_w = next_w;
+ continue;
- (StgClosure *)tmp = isAlive((StgClosure *)t);
-
- if (tmp != NULL) {
- t = tmp;
- }
+ case WEAK:
+ /* Now, check whether the key is reachable.
+ */
+ new = isAlive(w->key);
+ if (new != NULL) {
+ w->key = new;
+ // evacuate the value and finalizer
+ w->value = evacuate(w->value);
+ w->finalizer = evacuate(w->finalizer);
+ // remove this weak ptr from the old_weak_ptr list
+ *last_w = w->link;
+ // and put it on the new weak ptr list
+ next_w = w->link;
+ w->link = weak_ptr_list;
+ weak_ptr_list = w;
+ flag = rtsTrue;
+ IF_DEBUG(weak, belch("Weak pointer still alive at %p -> %p",
+ w, w->key));
+ continue;
+ }
+ else {
+ last_w = &(w->link);
+ next_w = w->link;
+ continue;
+ }
- ASSERT(get_itbl(t)->type == TSO);
- switch (t->what_next) {
- case ThreadRelocated:
- next = t->link;
- *prev = next;
- continue;
- case ThreadKilled:
- case ThreadComplete:
- // finshed or died. The thread might still be alive, but we
- // don't keep it on the all_threads list. Don't forget to
- // stub out its global_link field.
- next = t->global_link;
- t->global_link = END_TSO_QUEUE;
- *prev = next;
- continue;
- default:
- ;
+ default:
+ barf("traverse_weak_ptr_list: not WEAK");
+ }
}
+
+ /* If we didn't make any changes, then we can go round and kill all
+ * the dead weak pointers. The old_weak_ptr list is used as a list
+ * of pending finalizers later on.
+ */
+ if (flag == rtsFalse) {
+ for (w = old_weak_ptr_list; w; w = w->link) {
+ w->finalizer = evacuate(w->finalizer);
+ }
- if (tmp == NULL) {
- // not alive (yet): leave this thread on the old_all_threads list.
- prev = &(t->global_link);
- next = t->global_link;
- }
- else {
- // alive: move this thread onto the all_threads list.
- next = t->global_link;
- t->global_link = all_threads;
- all_threads = t;
- *prev = next;
+ // Next, move to the WeakThreads stage after fully
+ // scavenging the finalizers we've just evacuated.
+ weak_stage = WeakThreads;
}
- }
- }
- /* If we didn't make any changes, then we can go round and kill all
- * the dead weak pointers. The old_weak_ptr list is used as a list
- * of pending finalizers later on.
- */
- if (flag == rtsFalse) {
- for (w = old_weak_ptr_list; w; w = w->link) {
- w->finalizer = evacuate(w->finalizer);
- }
+ return rtsTrue;
- /* And resurrect any threads which were about to become garbage.
- */
- {
- StgTSO *t, *tmp, *next;
- for (t = old_all_threads; t != END_TSO_QUEUE; t = next) {
- next = t->global_link;
- (StgClosure *)tmp = evacuate((StgClosure *)t);
- tmp->global_link = resurrected_threads;
- resurrected_threads = tmp;
+ case WeakThreads:
+ /* Now deal with the all_threads list, which behaves somewhat like
+ * the weak ptr list. If we discover any threads that are about to
+ * become garbage, we wake them up and administer an exception.
+ */
+ {
+ StgTSO *t, *tmp, *next, **prev;
+
+ prev = &old_all_threads;
+ for (t = old_all_threads; t != END_TSO_QUEUE; t = next) {
+
+ (StgClosure *)tmp = isAlive((StgClosure *)t);
+
+ if (tmp != NULL) {
+ t = tmp;
+ }
+
+ ASSERT(get_itbl(t)->type == TSO);
+ switch (t->what_next) {
+ case ThreadRelocated:
+ next = t->link;
+ *prev = next;
+ continue;
+ case ThreadKilled:
+ case ThreadComplete:
+ // finshed or died. The thread might still be alive, but we
+ // don't keep it on the all_threads list. Don't forget to
+ // stub out its global_link field.
+ next = t->global_link;
+ t->global_link = END_TSO_QUEUE;
+ *prev = next;
+ continue;
+ default:
+ ;
+ }
+
+ if (tmp == NULL) {
+ // not alive (yet): leave this thread on the
+ // old_all_threads list.
+ prev = &(t->global_link);
+ next = t->global_link;
+ }
+ else {
+ // alive: move this thread onto the all_threads list.
+ next = t->global_link;
+ t->global_link = all_threads;
+ all_threads = t;
+ *prev = next;
+ }
+ }
}
- }
+
+ /* And resurrect any threads which were about to become garbage.
+ */
+ {
+ StgTSO *t, *tmp, *next;
+ for (t = old_all_threads; t != END_TSO_QUEUE; t = next) {
+ next = t->global_link;
+ (StgClosure *)tmp = evacuate((StgClosure *)t);
+ tmp->global_link = resurrected_threads;
+ resurrected_threads = tmp;
+ }
+ }
+
+ weak_stage = WeakDone; // *now* we're done,
+ return rtsTrue; // but one more round of scavenging, please
- weak_done = rtsTrue;
+ default:
+ barf("traverse_weak_ptr_list");
}
- return rtsTrue;
}
/* -----------------------------------------------------------------------------
last_w = list;
for (w = *list; w; w = w->link) {
+ // w might be WEAK, EVACUATED, or DEAD_WEAK (actually CON_STATIC) here
+ ASSERT(w->header.info == &stg_DEAD_WEAK_info
+ || get_itbl(w)->type == WEAK || get_itbl(w)->type == EVACUATED);
(StgClosure *)w = evacuate((StgClosure *)w);
*last_w = w;
last_w = &(w->link);
while (1) {
+ ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
info = get_itbl(p);
- /* 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.
- */
+ // ignore static closures
+ //
+ // 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(p)) {
+ return p;
+ }
- loop:
- bd = Bdescr((P_)p);
// ignore closures in generations that we're not collecting.
- if (LOOKS_LIKE_STATIC(p) || bd->gen_no > N) {
+ bd = Bdescr((P_)p);
+ if (bd->gen_no > N) {
+ return p;
+ }
+
+ // if it's a pointer into to-space, then we're done
+ if (bd->flags & BF_EVACUATED) {
return p;
}
- // large objects have an evacuated flag
+
+ // large objects use the evacuated flag
if (bd->flags & BF_LARGE) {
- if (bd->flags & BF_EVACUATED) {
- return p;
- } else {
- return NULL;
- }
+ return NULL;
}
+
// check the mark bit for compacted steps
if (bd->step->is_compacted && is_marked((P_)p,bd)) {
return p;
case TSO:
if (((StgTSO *)p)->what_next == ThreadRelocated) {
p = (StgClosure *)((StgTSO *)p)->link;
- goto loop;
- }
+ continue;
+ }
+ return NULL;
default:
// dead.
*root = evacuate(*root);
}
-static void
-addBlock(step *stp)
-{
- bdescr *bd = allocBlock();
- bd->gen_no = stp->gen_no;
- bd->step = stp;
-
- if (stp->gen_no <= N) {
- bd->flags = BF_EVACUATED;
- } else {
- bd->flags = 0;
- }
-
- stp->hp_bd->free = stp->hp;
- stp->hp_bd->link = bd;
- stp->hp = bd->start;
- stp->hpLim = stp->hp + BLOCK_SIZE_W;
- stp->hp_bd = bd;
- stp->n_to_blocks++;
- new_blocks++;
-}
-
-
static __inline__ void
upd_evacuee(StgClosure *p, StgClosure *dest)
{
- p->header.info = &stg_EVACUATED_info;
- ((StgEvacuated *)p)->evacuee = dest;
+ // Source object must be in from-space:
+ ASSERT((Bdescr((P_)p)->flags & BF_EVACUATED) == 0);
+ // not true: (ToDo: perhaps it should be)
+ // ASSERT(Bdescr((P_)dest)->flags & BF_EVACUATED);
+ p->header.info = &stg_EVACUATED_info;
+ ((StgEvacuated *)p)->evacuee = dest;
}
copy(StgClosure *src, nat size, step *stp)
{
P_ to, from, dest;
+#ifdef PROFILING
+ // @LDV profiling
+ nat size_org = size;
+#endif
TICK_GC_WORDS_COPIED(size);
/* Find out where we're going, using the handy "to" pointer in
* necessary.
*/
if (stp->hp + size >= stp->hpLim) {
- addBlock(stp);
+ gc_alloc_block(stp);
}
for(to = stp->hp, from = (P_)src; size>0; --size) {
dest = stp->hp;
stp->hp = to;
upd_evacuee(src,(StgClosure *)dest);
+#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(src, size_org);
+#endif
return (StgClosure *)dest;
}
*/
-static __inline__ StgClosure *
+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;
+#endif
TICK_GC_WORDS_COPIED(size_to_copy);
if (stp->gen_no < evac_gen) {
}
if (stp->hp + size_to_reserve >= stp->hpLim) {
- addBlock(stp);
+ gc_alloc_block(stp);
}
for(to = stp->hp, from = (P_)src; size_to_copy>0; --size_to_copy) {
dest = stp->hp;
stp->hp += size_to_reserve;
upd_evacuee(src,(StgClosure *)dest);
+#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)
+ FILL_SLOP(stp->hp - 1, (int)(size_to_reserve - size_to_copy_org));
+#endif
return (StgClosure *)dest;
}
Evacuate a large object
This just consists of removing the object from the (doubly-linked)
- large_alloc_list, and linking it on to the (singly-linked)
- new_large_objects list, from where it will be scavenged later.
+ step->large_objects list, and linking it on to the (singly-linked)
+ step->new_large_objects list, from where it will be scavenged later.
Convention: bd->flags has BF_EVACUATED set for a large object
that has been evacuated, or unset otherwise.
bdescr *bd = Bdescr(p);
step *stp;
- // should point to the beginning of the block
- ASSERT(((W_)p & BLOCK_MASK) == 0);
-
+ // object must be at the beginning of the block (or be a ByteArray)
+ ASSERT(get_itbl((StgClosure *)p)->type == ARR_WORDS ||
+ (((W_)p & BLOCK_MASK) == 0));
+
// already evacuated?
if (bd->flags & BF_EVACUATED) {
/* Don't forget to set the failed_to_evac flag if we didn't get
the promotion until the next GC.
-------------------------------------------------------------------------- */
-
static StgClosure *
mkMutCons(StgClosure *ptr, generation *gen)
{
* necessary.
*/
if (stp->hp + sizeofW(StgIndOldGen) >= stp->hpLim) {
- addBlock(stp);
+ gc_alloc_block(stp);
}
q = (StgMutVar *)stp->hp;
#endif
// make sure the info pointer is into text space
- ASSERT(q && (LOOKS_LIKE_GHC_INFO(GET_INFO(q))
- || IS_HUGS_CONSTR_INFO(GET_INFO(q))));
+ ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
info = get_itbl(q);
switch (info -> type) {
case MUT_VAR:
case MVAR:
- to = copy(q,sizeW_fromITBL(info),stp);
- return to;
+ return copy(q,sizeW_fromITBL(info),stp);
case CONSTR_0_1:
{
case THUNK_SELECTOR:
{
- const StgInfoTable* selectee_info;
- StgClosure* selectee = ((StgSelector*)q)->selectee;
+ StgClosure *p;
- selector_loop:
- selectee_info = get_itbl(selectee);
- switch (selectee_info->type) {
- 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:
- {
- StgWord offset = info->layout.selector_offset;
-
- // check that the size is in range
- ASSERT(offset <
- (StgWord32)(selectee_info->layout.payload.ptrs +
- selectee_info->layout.payload.nptrs));
-
- // perform the selection!
- q = selectee->payload[offset];
-
- /* if we're already in to-space, there's no need to continue
- * with the evacuation, just update the source address with
- * a pointer to the (evacuated) constructor field.
- */
- if (HEAP_ALLOCED(q)) {
- bdescr *bd = Bdescr((P_)q);
- if (bd->flags & BF_EVACUATED) {
- if (bd->gen_no < evac_gen) {
- failed_to_evac = rtsTrue;
- TICK_GC_FAILED_PROMOTION();
- }
- return q;
- }
- }
-
- /* otherwise, carry on and evacuate this constructor field,
- * (but not the constructor itself)
- */
- goto loop;
+ if (thunk_selector_depth > MAX_THUNK_SELECTOR_DEPTH) {
+ return copy(q,THUNK_SELECTOR_sizeW(),stp);
}
- case IND:
- case IND_STATIC:
- case IND_PERM:
- case IND_OLDGEN:
- case IND_OLDGEN_PERM:
- selectee = ((StgInd *)selectee)->indirectee;
- goto selector_loop;
-
- case EVACUATED:
- selectee = ((StgEvacuated *)selectee)->evacuee;
- goto selector_loop;
-
- case THUNK_SELECTOR:
-# if 0
- /* Disabled 03 April 2001 by JRS; it seems to cause the GC (or
- something) to go into an infinite loop when the nightly
- stage2 compiles PrelTup.lhs. */
-
- /* we can't recurse indefinitely in evacuate(), so set a
- * limit on the number of times we can go around this
- * loop.
- */
- if (thunk_selector_depth < MAX_THUNK_SELECTOR_DEPTH) {
- bdescr *bd;
- bd = Bdescr((P_)selectee);
- if (!bd->flags & BF_EVACUATED) {
- thunk_selector_depth++;
- selectee = evacuate(selectee);
- thunk_selector_depth--;
- goto selector_loop;
- }
- }
- // otherwise, fall through...
-# endif
-
- case AP_UPD:
- 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:
- case CAF_BLACKHOLE:
- case SE_CAF_BLACKHOLE:
- case SE_BLACKHOLE:
- case BLACKHOLE:
- case BLACKHOLE_BQ:
- // not evaluated yet
- break;
-
-#if defined(PAR)
- // a copy of the top-level cases below
- case RBH: // cf. BLACKHOLE_BQ
- {
- //StgInfoTable *rip = get_closure_info(q, &size, &ptrs, &nonptrs, &vhs, str);
- to = copy(q,BLACKHOLE_sizeW(),stp);
- //ToDo: derive size etc from reverted IP
- //to = copy(q,size,stp);
- // recordMutable((StgMutClosure *)to);
- return to;
- }
-
- case BLOCKED_FETCH:
- ASSERT(sizeofW(StgBlockedFetch) >= MIN_NONUPD_SIZE);
- to = copy(q,sizeofW(StgBlockedFetch),stp);
- return to;
+ p = eval_thunk_selector(info->layout.selector_offset,
+ (StgSelector *)q);
-# ifdef DIST
- case REMOTE_REF:
-# endif
- case FETCH_ME:
- ASSERT(sizeofW(StgBlockedFetch) >= MIN_UPD_SIZE);
- to = copy(q,sizeofW(StgFetchMe),stp);
- return to;
-
- case FETCH_ME_BQ:
- ASSERT(sizeofW(StgBlockedFetch) >= MIN_UPD_SIZE);
- to = copy(q,sizeofW(StgFetchMeBlockingQueue),stp);
- return to;
+ if (p == NULL) {
+ return copy(q,THUNK_SELECTOR_sizeW(),stp);
+ } else {
+ // q is still BLACKHOLE'd.
+ thunk_selector_depth++;
+ p = evacuate(p);
+ thunk_selector_depth--;
+ upd_evacuee(q,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
-
- default:
- barf("evacuate: THUNK_SELECTOR: strange selectee %d",
- (int)(selectee_info->type));
- }
+ return p;
+ }
}
- return copy(q,THUNK_SELECTOR_sizeW(),stp);
case IND:
case IND_OLDGEN:
case UPDATE_FRAME:
case STOP_FRAME:
case CATCH_FRAME:
- case SEQ_FRAME:
// shouldn't see these
barf("evacuate: stack frame at %p\n", q);
- case AP_UPD:
case PAP:
- /* PAPs and AP_UPDs are special - the payload is a copy of a chunk
- * of stack, tagging and all.
- */
+ case AP:
return copy(q,pap_sizeW((StgPAP*)q),stp);
+ case AP_STACK:
+ return copy(q,ap_stack_sizeW((StgAP_STACK*)q),stp);
+
case EVACUATED:
/* Already evacuated, just return the forwarding address.
* HOWEVER: if the requested destination generation (evac_gen) is
*/
if (evac_gen > 0) { // optimisation
StgClosure *p = ((StgEvacuated*)q)->evacuee;
- if (Bdescr((P_)p)->gen_no < evac_gen) {
+ if (HEAP_ALLOCED(p) && Bdescr((P_)p)->gen_no < evac_gen) {
failed_to_evac = rtsTrue;
TICK_GC_FAILED_PROMOTION();
}
}
/* -----------------------------------------------------------------------------
- move_TSO is called to update the TSO structure after it has been
- moved from one place to another.
+ Evaluate a THUNK_SELECTOR if possible.
+
+ returns: NULL if we couldn't evaluate this THUNK_SELECTOR, or
+ a closure pointer if we evaluated it and this is the result. Note
+ that "evaluating" the THUNK_SELECTOR doesn't necessarily mean
+ reducing it to HNF, just that we have eliminated the selection.
+ The result might be another thunk, or even another THUNK_SELECTOR.
+
+ If the return value is non-NULL, the original selector thunk has
+ been BLACKHOLE'd, and should be updated with an indirection or a
+ forwarding pointer. If the return value is NULL, then the selector
+ thunk is unchanged.
-------------------------------------------------------------------------- */
-void
-move_TSO(StgTSO *src, StgTSO *dest)
+static StgClosure *
+eval_thunk_selector( nat field, StgSelector * p )
{
- ptrdiff_t diff;
+ StgInfoTable *info;
+ const StgInfoTable *info_ptr;
+ StgClosure *selectee;
+
+ selectee = p->selectee;
- // relocate the stack pointers...
- diff = (StgPtr)dest - (StgPtr)src; // In *words*
- dest->sp = (StgPtr)dest->sp + diff;
- dest->su = (StgUpdateFrame *) ((P_)dest->su + diff);
+ // Save the real info pointer (NOTE: not the same as get_itbl()).
+ info_ptr = p->header.info;
- relocate_stack(dest, diff);
-}
+ // If the THUNK_SELECTOR is in a generation that we are not
+ // collecting, then bail out early. We won't be able to save any
+ // space in any case, and updating with an indirection is trickier
+ // in an old gen.
+ if (Bdescr((StgPtr)p)->gen_no > N) {
+ return NULL;
+ }
-/* -----------------------------------------------------------------------------
- relocate_stack is called to update the linkage between
- UPDATE_FRAMEs (and SEQ_FRAMEs etc.) when a stack is moved from one
- place to another.
- -------------------------------------------------------------------------- */
+ // BLACKHOLE 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.
+ SET_INFO(p,&stg_BLACKHOLE_info);
-StgTSO *
-relocate_stack(StgTSO *dest, ptrdiff_t diff)
-{
- StgUpdateFrame *su;
- StgCatchFrame *cf;
- StgSeqFrame *sf;
+selector_loop:
- su = dest->su;
+ info = get_itbl(selectee);
+ switch (info->type) {
+ 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));
+
+ return selectee->payload[field];
- while ((P_)su < dest->stack + dest->stack_size) {
- switch (get_itbl(su)->type) {
-
- // GCC actually manages to common up these three cases!
+ case IND:
+ case IND_PERM:
+ case IND_OLDGEN:
+ case IND_OLDGEN_PERM:
+ selectee = ((StgInd *)selectee)->indirectee;
+ goto selector_loop;
- case UPDATE_FRAME:
- su->link = (StgUpdateFrame *) ((StgPtr)su->link + diff);
- su = su->link;
- continue;
+ case EVACUATED:
+ // 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.
+ break;
- case CATCH_FRAME:
- cf = (StgCatchFrame *)su;
- cf->link = (StgUpdateFrame *) ((StgPtr)cf->link + diff);
- su = cf->link;
- continue;
+ case IND_STATIC:
+ // We can't easily tell whether the indirectee is into
+ // from or to-space, so just bail out here.
+ break;
- case SEQ_FRAME:
- sf = (StgSeqFrame *)su;
- sf->link = (StgUpdateFrame *) ((StgPtr)sf->link + diff);
- su = sf->link;
- continue;
+ case THUNK_SELECTOR:
+ {
+ StgClosure *val;
- case STOP_FRAME:
- // all done!
- break;
+ // check that we don't recurse too much, re-using the
+ // depth bound also used in evacuate().
+ thunk_selector_depth++;
+ if (thunk_selector_depth > MAX_THUNK_SELECTOR_DEPTH) {
+ break;
+ }
- default:
- barf("relocate_stack %d", (int)(get_itbl(su)->type));
+ val = eval_thunk_selector(info->layout.selector_offset,
+ (StgSelector *)selectee);
+
+ thunk_selector_depth--;
+
+ if (val == NULL) {
+ break;
+ } else {
+ // We evaluated this selector thunk, so update it with
+ // an indirection. NOTE: we don't use UPD_IND here,
+ // because we are guaranteed that p is in a generation
+ // that we are collecting, and we never want to put the
+ // indirection on a mutable list.
+ ((StgInd *)selectee)->indirectee = val;
+ SET_INFO(selectee,&stg_IND_info);
+ selectee = val;
+ goto selector_loop;
+ }
+ }
+
+ case AP:
+ 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:
+ case CAF_BLACKHOLE:
+ case SE_CAF_BLACKHOLE:
+ case SE_BLACKHOLE:
+ case BLACKHOLE:
+ case BLACKHOLE_BQ:
+#if defined(PAR)
+ case RBH:
+ case BLOCKED_FETCH:
+# ifdef DIST
+ case REMOTE_REF:
+# endif
+ case FETCH_ME:
+ case FETCH_ME_BQ:
+#endif
+ // not evaluated yet
+ break;
+
+ default:
+ barf("eval_thunk_selector: strange selectee %d",
+ (int)(info->type));
}
- break;
- }
- return dest;
+ // We didn't manage to evaluate this thunk; restore the old info pointer
+ SET_INFO(p, info_ptr);
+ return NULL;
}
+/* -----------------------------------------------------------------------------
+ move_TSO is called to update the TSO structure after it has been
+ moved from one place to another.
+ -------------------------------------------------------------------------- */
+
+void
+move_TSO (StgTSO *src, StgTSO *dest)
+{
+ ptrdiff_t diff;
+ // relocate the stack pointers...
+ diff = (StgPtr)dest - (StgPtr)src; // In *words*
+ dest->sp = (StgPtr)dest->sp + diff;
+}
+/* evacuate the SRT. If srt_len is zero, then there isn't an
+ * srt field in the info table. That's ok, because we'll
+ * never dereference it.
+ */
static inline void
-scavenge_srt(const StgInfoTable *info)
+scavenge_srt (StgClosure **srt, nat srt_len)
{
- StgClosure **srt, **srt_end;
+ StgClosure **srt_end;
+
+ srt_end = srt + srt_len;
- /* evacuate the SRT. If srt_len is zero, then there isn't an
- * srt field in the info table. That's ok, because we'll
- * never dereference it.
- */
- srt = (StgClosure **)(info->srt);
- srt_end = srt + info->srt_len;
for (; srt < srt_end; srt++) {
/* Special-case to handle references to closures hiding out in DLLs, since
double indirections required to get at those. The code generator knows
}
}
+
+static inline void
+scavenge_thunk_srt(const StgInfoTable *info)
+{
+ StgThunkInfoTable *thunk_info;
+
+ thunk_info = itbl_to_thunk_itbl(info);
+ scavenge_srt((StgClosure **)thunk_info->srt, thunk_info->i.srt_len);
+}
+
+static inline void
+scavenge_fun_srt(const StgInfoTable *info)
+{
+ StgFunInfoTable *fun_info;
+
+ fun_info = itbl_to_fun_itbl(info);
+ scavenge_srt((StgClosure **)fun_info->srt, fun_info->i.srt_len);
+}
+
+static inline void
+scavenge_ret_srt(const StgInfoTable *info)
+{
+ StgRetInfoTable *ret_info;
+
+ ret_info = itbl_to_ret_itbl(info);
+ scavenge_srt((StgClosure **)ret_info->srt, ret_info->i.srt_len);
+}
+
/* -----------------------------------------------------------------------------
Scavenge a TSO.
-------------------------------------------------------------------------- */
static void
scavengeTSO (StgTSO *tso)
{
- // chase the link field for any TSOs on the same queue
- (StgClosure *)tso->link = evacuate((StgClosure *)tso->link);
- if ( tso->why_blocked == BlockedOnMVar
- || tso->why_blocked == BlockedOnBlackHole
- || tso->why_blocked == BlockedOnException
+ // chase the link field for any TSOs on the same queue
+ (StgClosure *)tso->link = evacuate((StgClosure *)tso->link);
+ if ( tso->why_blocked == BlockedOnMVar
+ || tso->why_blocked == BlockedOnBlackHole
+ || tso->why_blocked == BlockedOnException
#if defined(PAR)
- || tso->why_blocked == BlockedOnGA
- || tso->why_blocked == BlockedOnGA_NoSend
+ || tso->why_blocked == BlockedOnGA
+ || tso->why_blocked == BlockedOnGA_NoSend
#endif
- ) {
- tso->block_info.closure = evacuate(tso->block_info.closure);
- }
- if ( tso->blocked_exceptions != NULL ) {
- tso->blocked_exceptions =
- (StgTSO *)evacuate((StgClosure *)tso->blocked_exceptions);
- }
- // scavenge this thread's stack
- scavenge_stack(tso->sp, &(tso->stack[tso->stack_size]));
+ ) {
+ tso->block_info.closure = evacuate(tso->block_info.closure);
+ }
+ if ( tso->blocked_exceptions != NULL ) {
+ tso->blocked_exceptions =
+ (StgTSO *)evacuate((StgClosure *)tso->blocked_exceptions);
+ }
+
+ // scavenge this thread's stack
+ scavenge_stack(tso->sp, &(tso->stack[tso->stack_size]));
+}
+
+/* -----------------------------------------------------------------------------
+ Blocks of function args occur on the stack (at the top) and
+ in PAPs.
+ -------------------------------------------------------------------------- */
+
+static inline StgPtr
+scavenge_arg_block (StgFunInfoTable *fun_info, StgClosure **args)
+{
+ StgPtr p;
+ StgWord bitmap;
+ nat size;
+
+ p = (StgPtr)args;
+ switch (fun_info->fun_type) {
+ case ARG_GEN:
+ bitmap = BITMAP_BITS(fun_info->bitmap);
+ size = BITMAP_SIZE(fun_info->bitmap);
+ goto small_bitmap;
+ case ARG_GEN_BIG:
+ size = ((StgLargeBitmap *)fun_info->bitmap)->size;
+ scavenge_large_bitmap(p, (StgLargeBitmap *)fun_info->bitmap, size);
+ p += size;
+ break;
+ default:
+ bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->fun_type]);
+ size = BITMAP_SIZE(stg_arg_bitmaps[fun_info->fun_type]);
+ small_bitmap:
+ while (size > 0) {
+ if ((bitmap & 1) == 0) {
+ (StgClosure *)*p = evacuate((StgClosure *)*p);
+ }
+ p++;
+ bitmap = bitmap >> 1;
+ size--;
+ }
+ break;
+ }
+ return p;
+}
+
+static inline StgPtr
+scavenge_PAP (StgPAP *pap)
+{
+ StgPtr p;
+ StgWord bitmap, size;
+ StgFunInfoTable *fun_info;
+
+ pap->fun = evacuate(pap->fun);
+ fun_info = get_fun_itbl(pap->fun);
+ ASSERT(fun_info->i.type != PAP);
+
+ p = (StgPtr)pap->payload;
+ size = pap->n_args;
+
+ switch (fun_info->fun_type) {
+ case ARG_GEN:
+ bitmap = BITMAP_BITS(fun_info->bitmap);
+ goto small_bitmap;
+ case ARG_GEN_BIG:
+ scavenge_large_bitmap(p, (StgLargeBitmap *)fun_info->bitmap, size);
+ p += size;
+ break;
+ case ARG_BCO:
+ scavenge_large_bitmap((StgPtr)pap->payload, BCO_BITMAP(pap->fun), size);
+ p += size;
+ break;
+ default:
+ bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->fun_type]);
+ small_bitmap:
+ size = pap->n_args;
+ while (size > 0) {
+ if ((bitmap & 1) == 0) {
+ (StgClosure *)*p = evacuate((StgClosure *)*p);
+ }
+ p++;
+ bitmap = bitmap >> 1;
+ size--;
+ }
+ break;
+ }
+ return p;
}
/* -----------------------------------------------------------------------------
continue;
}
+ ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
info = get_itbl((StgClosure *)p);
- ASSERT(p && (LOOKS_LIKE_GHC_INFO(info) || IS_HUGS_CONSTR_INFO(info)));
+ ASSERT(thunk_selector_depth == 0);
+
q = p;
switch (info->type) {
break;
}
- case THUNK_2_0:
case FUN_2_0:
- scavenge_srt(info);
+ scavenge_fun_srt(info);
+ ((StgClosure *)p)->payload[1] = evacuate(((StgClosure *)p)->payload[1]);
+ ((StgClosure *)p)->payload[0] = evacuate(((StgClosure *)p)->payload[0]);
+ p += sizeofW(StgHeader) + 2;
+ break;
+
+ case THUNK_2_0:
+ scavenge_thunk_srt(info);
case CONSTR_2_0:
((StgClosure *)p)->payload[1] = evacuate(((StgClosure *)p)->payload[1]);
((StgClosure *)p)->payload[0] = evacuate(((StgClosure *)p)->payload[0]);
break;
case THUNK_1_0:
- scavenge_srt(info);
+ scavenge_thunk_srt(info);
((StgClosure *)p)->payload[0] = evacuate(((StgClosure *)p)->payload[0]);
p += sizeofW(StgHeader) + 2; // MIN_UPD_SIZE
break;
case FUN_1_0:
- scavenge_srt(info);
+ scavenge_fun_srt(info);
case CONSTR_1_0:
((StgClosure *)p)->payload[0] = evacuate(((StgClosure *)p)->payload[0]);
p += sizeofW(StgHeader) + 1;
break;
case THUNK_0_1:
- scavenge_srt(info);
+ scavenge_thunk_srt(info);
p += sizeofW(StgHeader) + 2; // MIN_UPD_SIZE
break;
case FUN_0_1:
- scavenge_srt(info);
+ scavenge_fun_srt(info);
case CONSTR_0_1:
p += sizeofW(StgHeader) + 1;
break;
case THUNK_0_2:
+ scavenge_thunk_srt(info);
+ p += sizeofW(StgHeader) + 2;
+ break;
+
case FUN_0_2:
- scavenge_srt(info);
+ scavenge_fun_srt(info);
case CONSTR_0_2:
p += sizeofW(StgHeader) + 2;
break;
case THUNK_1_1:
+ scavenge_thunk_srt(info);
+ ((StgClosure *)p)->payload[0] = evacuate(((StgClosure *)p)->payload[0]);
+ p += sizeofW(StgHeader) + 2;
+ break;
+
case FUN_1_1:
- scavenge_srt(info);
+ scavenge_fun_srt(info);
case CONSTR_1_1:
((StgClosure *)p)->payload[0] = evacuate(((StgClosure *)p)->payload[0]);
p += sizeofW(StgHeader) + 2;
break;
case FUN:
+ scavenge_fun_srt(info);
+ goto gen_obj;
+
case THUNK:
- scavenge_srt(info);
+ scavenge_thunk_srt(info);
// fall through
+ gen_obj:
case CONSTR:
case WEAK:
case FOREIGN:
}
case IND_PERM:
- if (stp->gen_no != 0) {
- SET_INFO(((StgClosure *)p), &stg_IND_OLDGEN_PERM_info);
- }
+ if (stp->gen->no != 0) {
+#ifdef PROFILING
+ // @LDV profiling
+ // No need to call LDV_recordDead_FILL_SLOP_DYNAMIC() because an
+ // IND_OLDGEN_PERM closure is larger than an IND_PERM closure.
+ LDV_recordDead((StgClosure *)p, sizeofW(StgInd));
+#endif
+ //
+ // Todo: maybe use SET_HDR() and remove LDV_recordCreate()?
+ //
+ SET_INFO(((StgClosure *)p), &stg_IND_OLDGEN_PERM_info);
+#ifdef PROFILING
+ // @LDV profiling
+ // We pretend that p has just been created.
+ LDV_recordCreate((StgClosure *)p);
+#endif
+ }
// fall through
case IND_OLDGEN_PERM:
((StgIndOldGen *)p)->indirectee =
break;
}
- case AP_UPD: // same as PAPs
- case PAP:
- /* Treat a PAP just like a section of stack, not forgetting to
- * evacuate the function pointer too...
- */
- {
- StgPAP* pap = (StgPAP *)p;
+ // A chunk of stack saved in a heap object
+ case AP_STACK:
+ {
+ StgAP_STACK *ap = (StgAP_STACK *)p;
- pap->fun = evacuate(pap->fun);
- scavenge_stack((P_)pap->payload, (P_)pap->payload + pap->n_args);
- p += pap_sizeW(pap);
+ ap->fun = evacuate(ap->fun);
+ scavenge_stack((StgPtr)ap->payload, (StgPtr)ap->payload + ap->size);
+ p = (StgPtr)ap->payload + ap->size;
break;
}
-
+
+ case PAP:
+ case AP:
+ p = scavenge_PAP((StgPAP *)p);
+ break;
+
case ARR_WORDS:
// nothing to follow
p += arr_words_sizeW((StgArrWords *)p);
while (!mark_stack_empty()) {
p = pop_mark_stack();
+ ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
info = get_itbl((StgClosure *)p);
- ASSERT(p && (LOOKS_LIKE_GHC_INFO(info) || IS_HUGS_CONSTR_INFO(info)));
q = p;
switch (info->type) {
}
case FUN_2_0:
+ scavenge_fun_srt(info);
+ ((StgClosure *)p)->payload[1] = evacuate(((StgClosure *)p)->payload[1]);
+ ((StgClosure *)p)->payload[0] = evacuate(((StgClosure *)p)->payload[0]);
+ break;
+
case THUNK_2_0:
- scavenge_srt(info);
+ scavenge_thunk_srt(info);
case CONSTR_2_0:
((StgClosure *)p)->payload[1] = evacuate(((StgClosure *)p)->payload[1]);
((StgClosure *)p)->payload[0] = evacuate(((StgClosure *)p)->payload[0]);
case FUN_1_0:
case FUN_1_1:
+ scavenge_fun_srt(info);
+ ((StgClosure *)p)->payload[0] = evacuate(((StgClosure *)p)->payload[0]);
+ break;
+
case THUNK_1_0:
case THUNK_1_1:
- scavenge_srt(info);
+ scavenge_thunk_srt(info);
case CONSTR_1_0:
case CONSTR_1_1:
((StgClosure *)p)->payload[0] = evacuate(((StgClosure *)p)->payload[0]);
case FUN_0_1:
case FUN_0_2:
+ scavenge_fun_srt(info);
+ break;
+
case THUNK_0_1:
case THUNK_0_2:
- scavenge_srt(info);
+ scavenge_thunk_srt(info);
+ break;
+
case CONSTR_0_1:
case CONSTR_0_2:
break;
case FUN:
+ scavenge_fun_srt(info);
+ goto gen_obj;
+
case THUNK:
- scavenge_srt(info);
+ scavenge_thunk_srt(info);
// fall through
+ gen_obj:
case CONSTR:
case WEAK:
case FOREIGN:
break;
}
- case AP_UPD: // same as PAPs
- case PAP:
- /* Treat a PAP just like a section of stack, not forgetting to
- * evacuate the function pointer too...
- */
- {
- StgPAP* pap = (StgPAP *)p;
+ // A chunk of stack saved in a heap object
+ case AP_STACK:
+ {
+ StgAP_STACK *ap = (StgAP_STACK *)p;
- pap->fun = evacuate(pap->fun);
- scavenge_stack((P_)pap->payload, (P_)pap->payload + pap->n_args);
+ ap->fun = evacuate(ap->fun);
+ scavenge_stack((StgPtr)ap->payload, (StgPtr)ap->payload + ap->size);
break;
}
+
+ case PAP:
+ case AP:
+ scavenge_PAP((StgPAP *)p);
+ break;
case MUT_ARR_PTRS:
// follow everything
nat saved_evac_gen = evac_gen;
rtsBool no_luck;
- ASSERT(p && (LOOKS_LIKE_GHC_INFO(GET_INFO((StgClosure *)p))
- || IS_HUGS_CONSTR_INFO(GET_INFO((StgClosure *)p))));
-
+ ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
info = get_itbl((StgClosure *)p);
switch (info->type) {
case ARR_WORDS:
// nothing to follow
break;
-
+
case MUT_ARR_PTRS:
{
// follow everything
break;
}
- case AP_UPD:
- case PAP:
- {
- StgPAP* pap = (StgPAP *)p;
- pap->fun = evacuate(pap->fun);
- scavenge_stack((P_)pap->payload, (P_)pap->payload + pap->n_args);
+ case AP_STACK:
+ {
+ StgAP_STACK *ap = (StgAP_STACK *)p;
+
+ ap->fun = evacuate(ap->fun);
+ scavenge_stack((StgPtr)ap->payload, (StgPtr)ap->payload + ap->size);
+ p = (StgPtr)ap->payload + ap->size;
break;
}
+ case PAP:
+ case AP:
+ p = scavenge_PAP((StgPAP *)p);
+ break;
+
case IND_OLDGEN:
// This might happen if for instance a MUT_CONS was pointing to a
// THUNK which has since been updated. The IND_OLDGEN will
for (; p != END_MUT_LIST; p = next, next = p->mut_link) {
- // make sure the info pointer is into text space
- ASSERT(p && (LOOKS_LIKE_GHC_INFO(GET_INFO(p))
- || IS_HUGS_CONSTR_INFO(GET_INFO(p))));
-
+ ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
info = get_itbl(p);
/*
if (info->type==RBH)
for (; p != END_MUT_LIST; p = next, next = p->mut_link) {
- // make sure the info pointer is into text space
- ASSERT(p && (LOOKS_LIKE_GHC_INFO(GET_INFO(p))
- || IS_HUGS_CONSTR_INFO(GET_INFO(p))));
-
+ ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
info = get_itbl(p);
/*
if (info->type==RBH)
list... */
while (p != END_OF_STATIC_LIST) {
+ ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
info = get_itbl(p);
/*
if (info->type==RBH)
info = REVERT_INFOPTR(info); // if it's an RBH, look at the orig closure
*/
// make sure the info pointer is into text space
- ASSERT(p && (LOOKS_LIKE_GHC_INFO(GET_INFO(p))
- || IS_HUGS_CONSTR_INFO(GET_INFO(p))));
/* Take this object *off* the static_objects list,
* and put it on the scavenged_static_objects list.
}
case THUNK_STATIC:
+ scavenge_thunk_srt(info);
+ break;
+
case FUN_STATIC:
- scavenge_srt(info);
+ scavenge_fun_srt(info);
break;
case CONSTR_STATIC:
}
/* -----------------------------------------------------------------------------
+ scavenge a chunk of memory described by a bitmap
+ -------------------------------------------------------------------------- */
+
+static void
+scavenge_large_bitmap( StgPtr p, StgLargeBitmap *large_bitmap, nat size )
+{
+ nat i, b;
+ StgWord bitmap;
+
+ b = 0;
+ bitmap = large_bitmap->bitmap[b];
+ for (i = 0; i < size; ) {
+ if ((bitmap & 1) == 0) {
+ (StgClosure *)*p = evacuate((StgClosure *)*p);
+ }
+ i++;
+ p++;
+ if (i % BITS_IN(W_) == 0) {
+ b++;
+ bitmap = large_bitmap->bitmap[b];
+ } else {
+ bitmap = bitmap >> 1;
+ }
+ }
+}
+
+static inline StgPtr
+scavenge_small_bitmap (StgPtr p, nat size, StgWord bitmap)
+{
+ while (size > 0) {
+ if ((bitmap & 1) == 0) {
+ (StgClosure *)*p = evacuate((StgClosure *)*p);
+ }
+ p++;
+ bitmap = bitmap >> 1;
+ size--;
+ }
+ return p;
+}
+
+/* -----------------------------------------------------------------------------
scavenge_stack walks over a section of stack and evacuates all the
objects pointed to by it. We can use the same code for walking
- PAPs, since these are just sections of copied stack.
+ AP_STACK_UPDs, since these are just sections of copied stack.
-------------------------------------------------------------------------- */
+
static void
scavenge_stack(StgPtr p, StgPtr stack_end)
{
- StgPtr q;
- const StgInfoTable* info;
+ const StgRetInfoTable* info;
StgWord bitmap;
+ nat size;
//IF_DEBUG(sanity, belch(" scavenging stack between %p and %p", p, stack_end));
/*
* Each time around this loop, we are looking at a chunk of stack
- * that starts with either a pending argument section or an
- * activation record.
+ * that starts with an activation record.
*/
while (p < stack_end) {
- q = *(P_ *)p;
-
- // If we've got a tag, skip over that many words on the stack
- if (IS_ARG_TAG((W_)q)) {
- p += ARG_SIZE(q);
- p++; continue;
- }
-
- /* Is q a pointer to a closure?
- */
- if (! LOOKS_LIKE_GHC_INFO(q) ) {
-#ifdef DEBUG
- if ( 0 && LOOKS_LIKE_STATIC_CLOSURE(q) ) { // Is it a static closure?
- ASSERT(closure_STATIC((StgClosure *)q));
- }
- // otherwise, must be a pointer into the allocation space.
-#endif
-
- (StgClosure *)*p = evacuate((StgClosure *)q);
- p++;
- continue;
- }
+ info = get_ret_itbl((StgClosure *)p);
- /*
- * Otherwise, q must be the info pointer of an activation
- * record. All activation records have 'bitmap' style layout
- * info.
- */
- info = get_itbl((StgClosure *)p);
-
- switch (info->type) {
+ switch (info->i.type) {
- // Dynamic bitmap: the mask is stored on the stack
- case RET_DYN:
- bitmap = ((StgRetDyn *)p)->liveness;
- p = (P_)&((StgRetDyn *)p)->payload[0];
- goto small_bitmap;
-
- // probably a slow-entry point return address:
- case FUN:
- case FUN_STATIC:
- {
-#if 0
- StgPtr old_p = p;
- p++; p++;
- IF_DEBUG(sanity,
- belch("HWL: scavenge_stack: FUN(_STATIC) adjusting p from %p to %p (instead of %p)",
- old_p, p, old_p+1));
-#else
- p++; // what if FHS!=1 !? -- HWL
-#endif
- goto follow_srt;
- }
-
- /* Specialised code for update frames, since they're so common.
- * We *know* the updatee points to a BLACKHOLE, CAF_BLACKHOLE,
- * or BLACKHOLE_BQ, so just inline the code to evacuate it here.
- */
case UPDATE_FRAME:
- {
- StgUpdateFrame *frame = (StgUpdateFrame *)p;
-
+ ((StgUpdateFrame *)p)->updatee
+ = evacuate(((StgUpdateFrame *)p)->updatee);
p += sizeofW(StgUpdateFrame);
-
-#ifndef not_yet
- frame->updatee = evacuate(frame->updatee);
continue;
-#else // specialised code for update frames, not sure if it's worth it.
- StgClosure *to;
- nat type = get_itbl(frame->updatee)->type;
-
- if (type == EVACUATED) {
- frame->updatee = evacuate(frame->updatee);
- continue;
- } else {
- bdescr *bd = Bdescr((P_)frame->updatee);
- step *stp;
- if (bd->gen_no > N) {
- if (bd->gen_no < evac_gen) {
- failed_to_evac = rtsTrue;
- }
- continue;
- }
-
- // Don't promote blackholes
- stp = bd->step;
- if (!(stp->gen_no == 0 &&
- stp->no != 0 &&
- stp->no == stp->gen->n_steps-1)) {
- stp = stp->to;
- }
-
- switch (type) {
- case BLACKHOLE:
- case CAF_BLACKHOLE:
- to = copyPart(frame->updatee, BLACKHOLE_sizeW(),
- sizeofW(StgHeader), stp);
- frame->updatee = to;
- continue;
- case BLACKHOLE_BQ:
- to = copy(frame->updatee, BLACKHOLE_sizeW(), stp);
- frame->updatee = to;
- recordMutable((StgMutClosure *)to);
- continue;
- default:
- /* will never be SE_{,CAF_}BLACKHOLE, since we
- don't push an update frame for single-entry thunks. KSW 1999-01. */
- barf("scavenge_stack: UPDATE_FRAME updatee");
- }
- }
-#endif
- }
// small bitmap (< 32 entries, or 64 on a 64-bit machine)
case STOP_FRAME:
case CATCH_FRAME:
- case SEQ_FRAME:
- case RET_BCO:
case RET_SMALL:
case RET_VEC_SMALL:
- bitmap = info->layout.bitmap;
- p++;
- // this assumes that the payload starts immediately after the info-ptr
- small_bitmap:
- while (bitmap != 0) {
- if ((bitmap & 1) == 0) {
- (StgClosure *)*p = evacuate((StgClosure *)*p);
- }
+ 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++;
- bitmap = bitmap >> 1;
- }
-
+ p = scavenge_small_bitmap(p, size, bitmap);
+
follow_srt:
- scavenge_srt(info);
- continue;
+ scavenge_srt((StgClosure **)info->srt, info->i.srt_len);
+ continue;
+
+ case RET_BCO: {
+ StgBCO *bco;
+ nat size;
+
+ p++;
+ (StgClosure *)*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:
case RET_VEC_BIG:
- {
- StgPtr q;
- StgLargeBitmap *large_bitmap;
- nat i;
+ {
+ nat size;
- large_bitmap = info->layout.large_bitmap;
+ size = info->i.layout.large_bitmap->size;
p++;
+ scavenge_large_bitmap(p, info->i.layout.large_bitmap, size);
+ p += size;
+ // and don't forget to follow the SRT
+ goto follow_srt;
+ }
- for (i=0; i<large_bitmap->size; i++) {
- bitmap = large_bitmap->bitmap[i];
- q = p + BITS_IN(W_);
- while (bitmap != 0) {
- if ((bitmap & 1) == 0) {
- (StgClosure *)*p = evacuate((StgClosure *)*p);
- }
+ // 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 = GET_LIVENESS(dyn);
+ p = (P_)&((StgRetDyn *)p)->payload[0];
+ size = RET_DYN_SIZE;
+ p = scavenge_small_bitmap(p, size, bitmap);
+
+ // skip over the non-ptr words
+ p += GET_NONPTRS(dyn);
+
+ // follow the ptr words
+ for (size = GET_PTRS(dyn); size > 0; size--) {
+ (StgClosure *)*p = evacuate((StgClosure *)*p);
p++;
- bitmap = bitmap >> 1;
- }
- if (i+1 < large_bitmap->size) {
- while (p < q) {
- (StgClosure *)*p = evacuate((StgClosure *)*p);
- p++;
- }
- }
}
+ continue;
+ }
- // and don't forget to follow the SRT
+ case RET_FUN:
+ {
+ StgRetFun *ret_fun = (StgRetFun *)p;
+ StgFunInfoTable *fun_info;
+
+ ret_fun->fun = evacuate(ret_fun->fun);
+ fun_info = get_fun_itbl(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->type));
+ barf("scavenge_stack: weird activation record found on stack: %d", (int)(info->i.type));
}
- }
+ }
}
/*-----------------------------------------------------------------------------
}
void
-scavengeCAFs( void )
+markCAFs( evac_fn evac )
{
StgIndStatic *c;
- evac_gen = 0;
for (c = (StgIndStatic *)caf_list; c != NULL;
c = (StgIndStatic *)c->static_link)
{
- c->indirectee = evacuate(c->indirectee);
+ evac(&c->indirectee);
}
}
static void
threadLazyBlackHole(StgTSO *tso)
{
- StgUpdateFrame *update_frame;
- StgBlockingQueue *bh;
- StgPtr stack_end;
-
- stack_end = &tso->stack[tso->stack_size];
- update_frame = tso->su;
-
- while (1) {
- switch (get_itbl(update_frame)->type) {
-
- case CATCH_FRAME:
- update_frame = ((StgCatchFrame *)update_frame)->link;
- break;
-
- case UPDATE_FRAME:
- bh = (StgBlockingQueue *)update_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;
- }
+ StgClosure *frame;
+ StgRetInfoTable *info;
+ StgBlockingQueue *bh;
+ StgPtr stack_end;
+
+ stack_end = &tso->stack[tso->stack_size];
+
+ frame = (StgClosure *)tso->sp;
- if (bh->header.info != &stg_BLACKHOLE_BQ_info &&
- bh->header.info != &stg_CAF_BLACKHOLE_info) {
+ while (1) {
+ info = get_ret_itbl(frame);
+
+ switch (info->i.type) {
+
+ case UPDATE_FRAME:
+ bh = (StgBlockingQueue *)((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_BLACKHOLE_BQ_info &&
+ bh->header.info != &stg_CAF_BLACKHOLE_info) {
#if (!defined(LAZY_BLACKHOLING)) && defined(DEBUG)
- belch("Unexpected lazy BHing required at 0x%04x",(int)bh);
+ belch("Unexpected lazy BHing required at 0x%04x",(int)bh);
#endif
- SET_INFO(bh,&stg_BLACKHOLE_info);
- }
-
- update_frame = update_frame->link;
- break;
-
- case SEQ_FRAME:
- update_frame = ((StgSeqFrame *)update_frame)->link;
- break;
-
- case STOP_FRAME:
- return;
- default:
- barf("threadPaused");
+#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);
+#ifdef PROFILING
+ // @LDV profiling
+ // We pretend that bh has just been created.
+ LDV_recordCreate(bh);
+#endif
+ }
+
+ frame = (StgClosure *) ((StgUpdateFrame *)frame + 1);
+ break;
+
+ case STOP_FRAME:
+ return;
+
+ // normal stack frames; do nothing except advance the pointer
+ default:
+ (StgPtr)frame += stack_frame_sizeW(frame);
+ }
}
- }
}
*
* -------------------------------------------------------------------------- */
+struct stack_gap { StgWord gap_size; struct stack_gap *next_gap; };
+
static void
threadSqueezeStack(StgTSO *tso)
{
- lnat displacement = 0;
- StgUpdateFrame *frame;
- StgUpdateFrame *next_frame; // Temporally next
- StgUpdateFrame *prev_frame; // Temporally previous
- StgPtr bottom;
- rtsBool prev_was_update_frame;
-#if DEBUG
- StgUpdateFrame *top_frame;
- nat upd_frames=0, stop_frames=0, catch_frames=0, seq_frames=0,
- bhs=0, squeezes=0;
- void printObj( StgClosure *obj ); // from Printer.c
+ StgPtr frame;
+ rtsBool prev_was_update_frame;
+ StgClosure *updatee = NULL;
+ StgPtr bottom;
+ StgRetInfoTable *info;
+ StgWord current_gap_size;
+ struct stack_gap *gap;
- top_frame = tso->su;
-#endif
-
- bottom = &(tso->stack[tso->stack_size]);
- frame = tso->su;
+ // Stage 1:
+ // Traverse the stack upwards, replacing adjacent update frames
+ // with a single update frame and a "stack gap". A stack gap
+ // contains two values: the size of the gap, and the distance
+ // to the next gap (or the stack top).
- /* There must be at least one frame, namely the STOP_FRAME.
- */
- ASSERT((P_)frame < bottom);
+ bottom = &(tso->stack[tso->stack_size]);
- /* Walk down the stack, reversing the links between frames so that
- * we can walk back up as we squeeze from the bottom. Note that
- * next_frame and prev_frame refer to next and previous as they were
- * added to the stack, rather than the way we see them in this
- * walk. (It makes the next loop less confusing.)
- *
- * Stop if we find an update frame pointing to a black hole
- * (see comment in threadLazyBlackHole()).
- */
-
- next_frame = NULL;
- // bottom - sizeof(StgStopFrame) is the STOP_FRAME
- while ((P_)frame < bottom - sizeofW(StgStopFrame)) {
- prev_frame = frame->link;
- frame->link = next_frame;
- next_frame = frame;
- frame = prev_frame;
-#if DEBUG
- IF_DEBUG(sanity,
- if (!(frame>=top_frame && frame<=(StgUpdateFrame *)bottom)) {
- printObj((StgClosure *)prev_frame);
- barf("threadSqueezeStack: current frame is rubbish %p; previous was %p\n",
- frame, prev_frame);
- })
- switch (get_itbl(frame)->type) {
- case UPDATE_FRAME:
- upd_frames++;
- if (frame->updatee->header.info == &stg_BLACKHOLE_info)
- bhs++;
- break;
- case STOP_FRAME:
- stop_frames++;
- break;
- case CATCH_FRAME:
- catch_frames++;
- break;
- case SEQ_FRAME:
- seq_frames++;
- break;
- default:
- barf("Found non-frame during stack squeezing at %p (prev frame was %p)\n",
- frame, prev_frame);
- printObj((StgClosure *)prev_frame);
- }
-#endif
- if (get_itbl(frame)->type == UPDATE_FRAME
- && frame->updatee->header.info == &stg_BLACKHOLE_info) {
- break;
- }
- }
+ frame = tso->sp;
- /* Now, we're at the bottom. Frame points to the lowest update
- * frame on the stack, and its link actually points to the frame
- * above. We have to walk back up the stack, squeezing out empty
- * update frames and turning the pointers back around on the way
- * back up.
- *
- * The bottom-most frame (the STOP_FRAME) has not been altered, and
- * we never want to eliminate it anyway. Just walk one step up
- * before starting to squeeze. When you get to the topmost frame,
- * remember that there are still some words above it that might have
- * to be moved.
- */
-
- prev_frame = frame;
- frame = next_frame;
+ ASSERT(frame < bottom);
+
+ prev_was_update_frame = rtsFalse;
+ current_gap_size = 0;
+ gap = (struct stack_gap *) (tso->sp - sizeofW(StgUpdateFrame));
- prev_was_update_frame = (get_itbl(prev_frame)->type == UPDATE_FRAME);
+ while (frame < bottom) {
+
+ info = get_ret_itbl((StgClosure *)frame);
+ switch (info->i.type) {
- /*
- * Loop through all of the frames (everything except the very
- * bottom). Things are complicated by the fact that we have
- * CATCH_FRAMEs and SEQ_FRAMEs interspersed with the update frames.
- * We can only squeeze when there are two consecutive UPDATE_FRAMEs.
- */
- while (frame != NULL) {
- StgPtr sp;
- StgPtr frame_bottom = (P_)frame + sizeofW(StgUpdateFrame);
- rtsBool is_update_frame;
-
- next_frame = frame->link;
- is_update_frame = (get_itbl(frame)->type == UPDATE_FRAME);
+ case UPDATE_FRAME:
+ {
+ StgUpdateFrame *upd = (StgUpdateFrame *)frame;
- /* Check to see if
- * 1. both the previous and current frame are update frames
- * 2. the current frame is empty
- */
- if (prev_was_update_frame && is_update_frame &&
- (P_)prev_frame == frame_bottom + displacement) {
-
- // Now squeeze out the current frame
- StgClosure *updatee_keep = prev_frame->updatee;
- StgClosure *updatee_bypass = frame->updatee;
-
-#if DEBUG
- IF_DEBUG(gc, belch("@@ squeezing frame at %p", frame));
- squeezes++;
-#endif
+ if (upd->updatee->header.info == &stg_BLACKHOLE_info) {
- /* Deal with blocking queues. If both updatees have blocked
- * threads, then we should merge the queues into the update
- * frame that we're keeping.
- *
- * Alternatively, we could just wake them up: they'll just go
- * straight to sleep on the proper blackhole! This is less code
- * and probably less bug prone, although it's probably much
- * slower --SDM
- */
-#if 0 // do it properly...
-# if (!defined(LAZY_BLACKHOLING)) && defined(DEBUG)
-# error Unimplemented lazy BH warning. (KSW 1999-01)
-# endif
- if (GET_INFO(updatee_bypass) == stg_BLACKHOLE_BQ_info
- || GET_INFO(updatee_bypass) == stg_CAF_BLACKHOLE_info
- ) {
- // Sigh. It has one. Don't lose those threads!
- if (GET_INFO(updatee_keep) == stg_BLACKHOLE_BQ_info) {
- // Urgh. Two queues. Merge them.
- P_ keep_tso = ((StgBlockingQueue *)updatee_keep)->blocking_queue;
-
- while (keep_tso->link != END_TSO_QUEUE) {
- keep_tso = keep_tso->link;
- }
- keep_tso->link = ((StgBlockingQueue *)updatee_bypass)->blocking_queue;
+ // found a BLACKHOLE'd update frame; we've been here
+ // before, in a previous GC, so just break out.
- } else {
- // For simplicity, just swap the BQ for the BH
- P_ temp = updatee_keep;
-
- updatee_keep = updatee_bypass;
- updatee_bypass = temp;
-
- // Record the swap in the kept frame (below)
- prev_frame->updatee = updatee_keep;
- }
- }
-#endif
+ // 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;
+ }
- TICK_UPD_SQUEEZED();
- /* wasn't there something about update squeezing and ticky to be
- * sorted out? oh yes: we aren't counting each enter properly
- * in this case. See the log somewhere. KSW 1999-04-21
- *
- * Check two things: that the two update frames don't point to
- * the same object, and that the updatee_bypass isn't already an
- * indirection. Both of these cases only happen when we're in a
- * block hole-style loop (and there are multiple update frames
- * on the stack pointing to the same closure), but they can both
- * screw us up if we don't check.
- */
- if (updatee_bypass != updatee_keep && !closure_IND(updatee_bypass)) {
- // this wakes the threads up
- UPD_IND_NOLOCK(updatee_bypass, updatee_keep);
- }
-
- sp = (P_)frame - 1; // sp = stuff to slide
- displacement += sizeofW(StgUpdateFrame);
-
- } else {
- // No squeeze for this frame
- sp = frame_bottom - 1; // Keep the current frame
-
- /* Do lazy black-holing.
- */
- if (is_update_frame) {
- StgBlockingQueue *bh = (StgBlockingQueue *)frame->updatee;
- if (bh->header.info != &stg_BLACKHOLE_info &&
- bh->header.info != &stg_BLACKHOLE_BQ_info &&
- bh->header.info != &stg_CAF_BLACKHOLE_info) {
+ if (prev_was_update_frame) {
+
+ TICK_UPD_SQUEEZED();
+ /* wasn't there something about update squeezing and ticky to be
+ * sorted out? oh yes: we aren't counting each enter properly
+ * in this case. See the log somewhere. KSW 1999-04-21
+ *
+ * Check two things: that the two update frames don't point to
+ * the same object, and that the updatee_bypass isn't already an
+ * indirection. Both of these cases only happen when we're in a
+ * block hole-style loop (and there are multiple update frames
+ * on the stack pointing to the same closure), but they can both
+ * screw us up if we don't check.
+ */
+ if (upd->updatee != updatee && !closure_IND(upd->updatee)) {
+ // this wakes the threads up
+ UPD_IND_NOLOCK(upd->updatee, updatee);
+ }
+
+ // now mark this update frame as a stack gap. The gap
+ // marker resides in the bottom-most update frame of
+ // the series of adjacent frames, and covers all the
+ // frames in this series.
+ current_gap_size += sizeofW(StgUpdateFrame);
+ ((struct stack_gap *)frame)->gap_size = current_gap_size;
+ ((struct stack_gap *)frame)->next_gap = gap;
+
+ frame += sizeofW(StgUpdateFrame);
+ continue;
+ }
+
+ // single update frame, or the topmost update frame in a series
+ else {
+ StgBlockingQueue *bh = (StgBlockingQueue *)upd->updatee;
+
+ // Do lazy black-holing
+ if (bh->header.info != &stg_BLACKHOLE_info &&
+ bh->header.info != &stg_BLACKHOLE_BQ_info &&
+ bh->header.info != &stg_CAF_BLACKHOLE_info) {
#if (!defined(LAZY_BLACKHOLING)) && defined(DEBUG)
- belch("Unexpected lazy BHing required at 0x%04x",(int)bh);
+ belch("Unexpected lazy BHing required at 0x%04x",(int)bh);
#endif
#ifdef DEBUG
- /* zero out the slop so that the sanity checker can tell
- * where the next closure is.
- */
- {
- StgInfoTable *info = get_itbl(bh);
- nat np = info->layout.payload.ptrs, nw = info->layout.payload.nptrs, i;
- /* don't zero out slop for a THUNK_SELECTOR, because its layout
- * info is used for a different purpose, and it's exactly the
- * same size as a BLACKHOLE in any case.
- */
- if (info->type != THUNK_SELECTOR) {
- for (i = np; i < np + nw; i++) {
- ((StgClosure *)bh)->payload[i] = 0;
- }
- }
- }
+ /* zero out the slop so that the sanity checker can tell
+ * where the next closure is.
+ */
+ {
+ StgInfoTable *bh_info = get_itbl(bh);
+ nat np = bh_info->layout.payload.ptrs,
+ nw = bh_info->layout.payload.nptrs, i;
+ /* don't zero out slop for a THUNK_SELECTOR,
+ * because its layout info is used for a
+ * different purpose, and it's exactly the
+ * same size as a BLACKHOLE in any case.
+ */
+ if (bh_info->type != THUNK_SELECTOR) {
+ for (i = np; i < np + nw; i++) {
+ ((StgClosure *)bh)->payload[i] = 0;
+ }
+ }
+ }
+#endif
+#ifdef PROFILING
+ // We pretend that bh is now dead.
+ LDV_recordDead_FILL_SLOP_DYNAMIC((StgClosure *)bh);
+#endif
+ // Todo: maybe use SET_HDR() and remove LDV_recordCreate()?
+ SET_INFO(bh,&stg_BLACKHOLE_info);
+#ifdef PROFILING
+ // We pretend that bh has just been created.
+ LDV_recordCreate(bh);
#endif
- SET_INFO(bh,&stg_BLACKHOLE_info);
+ }
+
+ prev_was_update_frame = rtsTrue;
+ updatee = upd->updatee;
+ frame += sizeofW(StgUpdateFrame);
+ continue;
+ }
}
- }
+
+ default:
+ prev_was_update_frame = rtsFalse;
- // Fix the link in the current frame (should point to the frame below)
- frame->link = prev_frame;
- prev_was_update_frame = is_update_frame;
- }
-
- // Now slide all words from sp up to the next frame
-
- if (displacement > 0) {
- P_ next_frame_bottom;
+ // we're not in a gap... check whether this is the end of a gap
+ // (an update frame can't be the end of a gap).
+ if (current_gap_size != 0) {
+ gap = (struct stack_gap *) (frame - sizeofW(StgUpdateFrame));
+ }
+ current_gap_size = 0;
- if (next_frame != NULL)
- next_frame_bottom = (P_)next_frame + sizeofW(StgUpdateFrame);
- else
- next_frame_bottom = tso->sp - 1;
-
-#if 0
- IF_DEBUG(gc,
- belch("sliding [%p, %p] by %ld", sp, next_frame_bottom,
- displacement))
-#endif
-
- while (sp >= next_frame_bottom) {
- sp[displacement] = *sp;
- sp -= 1;
- }
+ frame += stack_frame_sizeW((StgClosure *)frame);
+ continue;
+ }
}
- (P_)prev_frame = (P_)frame + displacement;
- frame = next_frame;
- }
- tso->sp += displacement;
- tso->su = prev_frame;
-#if 0
- IF_DEBUG(gc,
- belch("@@ threadSqueezeStack: squeezed %d update-frames; found %d BHs; found %d update-, %d stop-, %d catch, %d seq-frames",
- squeezes, bhs, upd_frames, stop_frames, catch_frames, seq_frames))
-#endif
-}
+done_traversing:
+
+ // 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:
+ //
+ // +| ********* |
+ // | ********* | <- sp
+ // | |
+ // | | <- gap_start
+ // | ......... | |
+ // | stack_gap | <- gap | chunk_size
+ // | ......... | |
+ // | ......... | <- gap_end v
+ // | ********* |
+ // | ********* |
+ // | ********* |
+ // -| ********* |
+ //
+ // 'sp' points the the current top-of-stack
+ // 'gap' points to the stack_gap structure inside the gap
+ // ***** indicates real stack data
+ // ..... indicates gap
+ // <empty> indicates unused
+ //
+ {
+ void *sp;
+ void *gap_start, *next_gap_start, *gap_end;
+ nat chunk_size;
+
+ next_gap_start = (void *)gap + sizeof(StgUpdateFrame);
+ sp = next_gap_start;
+ while ((StgPtr)gap > tso->sp) {
+
+ // we're working in *bytes* now...
+ gap_start = next_gap_start;
+ gap_end = gap_start - gap->gap_size * sizeof(W_);
+
+ gap = gap->next_gap;
+ next_gap_start = (void *)gap + sizeof(StgUpdateFrame);
+
+ chunk_size = gap_end - next_gap_start;
+ sp -= chunk_size;
+ memmove(sp, next_gap_start, chunk_size);
+ }
+
+ tso->sp = (StgPtr)sp;
+ }
+}
/* -----------------------------------------------------------------------------
* Pausing a thread