X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FGC.c;h=a5abcd95e92338ed8ec2b7458fee981b43e2c71b;hb=8d0e9847953007239e2c19dd1987f0a78825b897;hp=695e132b0c3f870cb3f28ebbae5cb01872f51c1f;hpb=6397979023625fb3d678642cb11ce709def12bc5;p=ghc-hetmet.git diff --git a/ghc/rts/GC.c b/ghc/rts/GC.c index 695e132..a5abcd9 100644 --- a/ghc/rts/GC.c +++ b/ghc/rts/GC.c @@ -1,5 +1,5 @@ /* ----------------------------------------------------------------------------- - * $Id: GC.c,v 1.7 1999/01/14 10:49:01 simonm Exp $ + * $Id: GC.c,v 1.23 1999/02/02 14:21:29 simonm Exp $ * * Two-space garbage collector * @@ -20,6 +20,7 @@ #include "DebugProf.h" #include "SchedAPI.h" #include "Weak.h" +#include "StablePriv.h" StgCAF* enteredCAFs; @@ -84,6 +85,10 @@ static rtsBool weak_done; /* all done for this pass */ */ static rtsBool failed_to_evac; +/* Old to-space (used for two-space collector only) + */ +bdescr *old_to_space; + /* ----------------------------------------------------------------------------- Static function declarations -------------------------------------------------------------------------- */ @@ -99,6 +104,7 @@ static void scavenge_large(step *step); static void scavenge(step *step); static void scavenge_static(void); static StgMutClosure *scavenge_mutable_list(StgMutClosure *p, nat gen); +static StgMutClosure *scavenge_mut_once_list(StgMutClosure *p, nat gen); #ifdef DEBUG static void gcCAFs(void); @@ -165,6 +171,7 @@ void GarbageCollect(void (*get_roots)(void)) /* Figure out which generation to collect */ + N = 0; for (g = 0; g < RtsFlags.GcFlags.generations; g++) { if (generations[g].steps[0].n_blocks >= generations[g].max_blocks) { N = g; @@ -185,18 +192,30 @@ void GarbageCollect(void (*get_roots)(void)) * zeroMutableList below). */ if (major_gc) { - zeroMutableList(generations[RtsFlags.GcFlags.generations-1].mut_list); + zeroMutableList(generations[RtsFlags.GcFlags.generations-1].mut_once_list); + } + + /* Save the old to-space if we're doing a two-space collection + */ + if (RtsFlags.GcFlags.generations == 1) { + old_to_space = g0s0->to_space; + g0s0->to_space = NULL; } /* 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; for (s = 0; s < generations[g].n_steps; s++) { + /* generation 0, step 0 doesn't need to-space */ - if (g == 0 && s == 0) { continue; } + if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) { + continue; + } + /* Get a free block for to-space. Extra blocks will be chained on * as necessary. */ @@ -250,21 +269,11 @@ void GarbageCollect(void (*get_roots)(void)) step->to_blocks = 0; step->new_large_objects = NULL; step->scavenged_large_objects = NULL; -#ifdef DEBUG - /* retain these so we can sanity-check later on */ - step->old_scan = step->scan; - step->old_scan_bd = step->scan_bd; -#endif } } /* ----------------------------------------------------------------------- - * follow all the roots that the application knows about. - */ - evac_gen = 0; - get_roots(); - - /* follow all the roots that we know about: + * follow all the roots that we know about: * - mutable lists from each generation > N * we want to *scavenge* these roots, not evacuate them: they're not * going to move in this GC. @@ -277,23 +286,32 @@ void GarbageCollect(void (*get_roots)(void)) */ { StgMutClosure *tmp, **pp; - for (g = RtsFlags.GcFlags.generations-1; g > N; g--) { - /* the act of scavenging the mutable list for this generation - * might place more objects on the mutable list itself. So we - * place the current mutable list in a temporary, scavenge it, - * and then append it to the new list. - */ - tmp = generations[g].mut_list; + for (g = N+1; g < RtsFlags.GcFlags.generations; g++) { + generations[g].saved_mut_list = generations[g].mut_list; generations[g].mut_list = END_MUT_LIST; - tmp = scavenge_mutable_list(tmp, g); + } + /* Do the mut-once lists first */ + for (g = RtsFlags.GcFlags.generations-1; g > N; g--) { + generations[g].mut_once_list = + scavenge_mut_once_list(generations[g].mut_once_list, g); + } + + for (g = RtsFlags.GcFlags.generations-1; g > N; g--) { + tmp = scavenge_mutable_list(generations[g].saved_mut_list, g); pp = &generations[g].mut_list; while (*pp != END_MUT_LIST) { pp = &(*pp)->mut_link; } *pp = tmp; } - } + } + + /* follow all the roots that the application knows about. + */ + evac_gen = 0; + get_roots(); + /* And don't forget to mark the TSO if we got here direct from * Haskell! */ if (CurrentTSO) { @@ -308,6 +326,10 @@ void GarbageCollect(void (*get_roots)(void)) weak_ptr_list = NULL; weak_done = rtsFalse; + /* Mark the stable pointer table. + */ + markStablePtrTable(major_gc); + #ifdef INTERPRETER { /* ToDo: To fix the caf leak, we need to make the commented out @@ -383,6 +405,35 @@ void GarbageCollect(void (*get_roots)(void)) } } + /* Now see which stable names are still alive + */ + gcStablePtrTable(major_gc); + + /* Set the maximum blocks for the oldest generation, based on twice + * the amount of live data now, adjusted to fit the maximum heap + * size if necessary. + * + * This is an approximation, since in the worst case we'll need + * twice the amount of live data plus whatever space the other + * generations need. + */ + if (RtsFlags.GcFlags.generations > 1) { + if (major_gc) { + oldest_gen->max_blocks = + stg_max(oldest_gen->steps[0].to_blocks * RtsFlags.GcFlags.oldGenFactor, + RtsFlags.GcFlags.minOldGenSize); + if (oldest_gen->max_blocks > RtsFlags.GcFlags.maxHeapSize / 2) { + oldest_gen->max_blocks = RtsFlags.GcFlags.maxHeapSize / 2; + if (((int)oldest_gen->max_blocks - + (int)oldest_gen->steps[0].to_blocks) < + (RtsFlags.GcFlags.pcFreeHeap * + RtsFlags.GcFlags.maxHeapSize / 200)) { + heapOverflow(); + } + } + } + } + /* run through all the generations/steps and tidy up */ for (g = 0; g < RtsFlags.GcFlags.generations; g++) { @@ -395,7 +446,7 @@ void GarbageCollect(void (*get_roots)(void)) bdescr *next; step = &generations[g].steps[s]; - if (!(g == 0 && s == 0)) { + if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) { /* Tidy the end of the to-space chains */ step->hp_bd->free = step->hp; step->hp_bd->link = NULL; @@ -436,24 +487,19 @@ void GarbageCollect(void (*get_roots)(void)) } step->large_objects = step->scavenged_large_objects; - /* Set the maximum blocks for this generation, - * using an arbitrary factor of the no. of blocks in step 0. + /* Set the maximum blocks for this generation, interpolating + * between the maximum size of the oldest and youngest + * generations. + * + * max_blocks = oldgen_max_blocks * G + * ---------------------- + * oldest_gen */ if (g != 0) { - generation *gen = &generations[g]; - gen->max_blocks = - stg_max(gen->steps[s].n_blocks * 2, - RtsFlags.GcFlags.minAllocAreaSize * 4); - if (gen->max_blocks > RtsFlags.GcFlags.maxHeapSize / 2) { - gen->max_blocks = RtsFlags.GcFlags.maxHeapSize / 2; - if (((int)gen->max_blocks - (int)gen->steps[0].n_blocks) < - (RtsFlags.GcFlags.pcFreeHeap * - RtsFlags.GcFlags.maxHeapSize / 200)) { - heapOverflow(); - } - } + generations[g].max_blocks = (oldest_gen->max_blocks * g) + / (RtsFlags.GcFlags.generations-1); } - + /* for older generations... */ } else { @@ -473,6 +519,100 @@ void GarbageCollect(void (*get_roots)(void)) } } + /* Guess the amount of live data for stats. */ + live = calcLive(); + + /* Two-space collector: + * Free the old to-space, and estimate the amount of live data. + */ + if (RtsFlags.GcFlags.generations == 1) { + nat blocks; + + if (old_to_space != NULL) { + freeChain(old_to_space); + } + for (bd = g0s0->to_space; bd != NULL; bd = bd->link) { + bd->evacuated = 0; /* now from-space */ + } + + /* For a two-space collector, we need to resize the nursery. */ + + /* set up a new nursery. Allocate a nursery size based on a + * function of the amount of live data (currently a factor of 2, + * should be configurable (ToDo)). Use the blocks from the old + * nursery if possible, freeing up any left over blocks. + * + * If we get near the maximum heap size, then adjust our nursery + * size accordingly. If the nursery is the same size as the live + * data (L), then we need 3L bytes. We can reduce the size of the + * nursery to bring the required memory down near 2L bytes. + * + * A normal 2-space collector would need 4L bytes to give the same + * performance we get from 3L bytes, reducing to the same + * performance at 2L bytes. + */ + blocks = g0s0->n_blocks; + + if ( blocks * RtsFlags.GcFlags.oldGenFactor * 2 > + RtsFlags.GcFlags.maxHeapSize ) { + int adjusted_blocks; /* signed on purpose */ + int pc_free; + + adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks); + IF_DEBUG(gc, fprintf(stderr, "Near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %d\n", RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks)); + pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize; + if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even be < 0 */ { + heapOverflow(); + } + blocks = adjusted_blocks; + + } else { + blocks *= RtsFlags.GcFlags.oldGenFactor; + if (blocks < RtsFlags.GcFlags.minAllocAreaSize) { + blocks = RtsFlags.GcFlags.minAllocAreaSize; + } + } + resizeNursery(blocks); + + } else { + /* Generational collector: + * If the user has given us a suggested heap size, adjust our + * allocation area to make best use of the memory available. + */ + + if (RtsFlags.GcFlags.heapSizeSuggestion) { + int blocks; + nat needed = calcNeeded(); /* approx blocks needed at next GC */ + + /* Guess how much will be live in generation 0 step 0 next time. + * A good approximation is the amount of data that was live this + * time: this assumes (1) that the size of G0S0 will be roughly + * the same as last time, and (2) that the promotion rate will be + * constant. + * + * If we don't know how much was live in G0S0 (because there's no + * step 1), then assume 30% (which is usually an overestimate). + */ + if (g0->n_steps == 1) { + needed += (g0s0->n_blocks * 30) / 100; + } else { + needed += g0->steps[1].n_blocks; + } + + /* Now we have a rough guess at the number of blocks needed for + * the next GC, subtract this from the user's suggested heap size + * and use the rest for the allocation area. + */ + blocks = (int)RtsFlags.GcFlags.heapSizeSuggestion - (int)needed; + + if (blocks < (int)RtsFlags.GcFlags.minAllocAreaSize) { + blocks = RtsFlags.GcFlags.minAllocAreaSize; + } + + resizeNursery((nat)blocks); + } + } + /* revert dead CAFs and update enteredCAFs list */ revertDeadCAFs(); @@ -495,19 +635,6 @@ void GarbageCollect(void (*get_roots)(void)) } current_nursery = g0s0->blocks; - live = 0; - for (g = 0; g < RtsFlags.GcFlags.generations; g++) { - for (s = 0; s < generations[g].n_steps; s++) { - /* approximate amount of live data (doesn't take into account slop - * at end of each block). ToDo: this more accurately. - */ - if (g == 0 && s == 0) { continue; } - step = &generations[g].steps[s]; - live += step->n_blocks * BLOCK_SIZE_W + - ((lnat)step->hp_bd->free -(lnat)step->hp_bd->start) / sizeof(W_); - } - } - /* Free the small objects allocated via allocate(), since this will * all have been copied into G0S1 now. */ @@ -522,24 +649,9 @@ void GarbageCollect(void (*get_roots)(void)) scheduleFinalisers(old_weak_ptr_list); /* check sanity after GC */ -#ifdef DEBUG - for (g = 0; g <= N; g++) { - for (s = 0; s < generations[g].n_steps; s++) { - if (g == 0 && s == 0) { continue; } - IF_DEBUG(sanity, checkHeap(generations[g].steps[s].blocks, NULL)); - IF_DEBUG(sanity, checkChain(generations[g].steps[s].large_objects)); - } - } - for (g = N+1; g < RtsFlags.GcFlags.generations; g++) { - for (s = 0; s < generations[g].n_steps; s++) { - IF_DEBUG(sanity, checkHeap(generations[g].steps[s].old_scan_bd, - generations[g].steps[s].old_scan)); - IF_DEBUG(sanity, checkChain(generations[g].steps[s].large_objects)); - } - } - IF_DEBUG(sanity, checkFreeListSanity()); -#endif + IF_DEBUG(sanity, checkSanity(N)); + /* extra GC trace info */ IF_DEBUG(gc, stat_describe_gens()); #ifdef DEBUG @@ -583,8 +695,7 @@ static rtsBool traverse_weak_ptr_list(void) { StgWeak *w, **last_w, *next_w; - StgClosure *target; - const StgInfoTable *info; + StgClosure *new; rtsBool flag = rtsFalse; if (weak_done) { return rtsFalse; } @@ -596,56 +707,26 @@ traverse_weak_ptr_list(void) last_w = &old_weak_ptr_list; for (w = old_weak_ptr_list; w; w = next_w) { - target = w->key; - loop: - /* ignore weak pointers in older generations */ - if (!LOOKS_LIKE_STATIC(target) && Bdescr((P_)target)->gen->no > N) { - IF_DEBUG(weak, fprintf(stderr,"Weak pointer still alive (in old gen) at %p\n", w)); - /* 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; - continue; - } - info = get_itbl(target); - switch (info->type) { - - case IND: - case IND_STATIC: - case IND_PERM: - case IND_OLDGEN: /* rely on compatible layout with StgInd */ - case IND_OLDGEN_PERM: - /* follow indirections */ - target = ((StgInd *)target)->indirectee; - goto loop; - - case EVACUATED: - /* If key is alive, evacuate value and finaliser and - * place weak ptr on new weak ptr list. - */ - IF_DEBUG(weak, fprintf(stderr,"Weak pointer still alive at %p\n", w)); - w->key = ((StgEvacuated *)target)->evacuee; + if ((new = isAlive(w->key))) { + w->key = new; + /* evacuate the value and finaliser */ w->value = evacuate(w->value); w->finaliser = evacuate(w->finaliser); - /* 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; - break; - - default: /* key is dead */ + IF_DEBUG(weak, fprintf(stderr,"Weak pointer still alive at %p -> %p\n", w, w->key)); + continue; + } + else { last_w = &(w->link); next_w = w->link; - break; + continue; } } @@ -664,14 +745,60 @@ traverse_weak_ptr_list(void) return rtsTrue; } +/* ----------------------------------------------------------------------------- + isAlive determines whether the given closure is still alive (after + a garbage collection) or not. It returns the new address of the + closure if it is alive, or NULL otherwise. + -------------------------------------------------------------------------- */ + +StgClosure * +isAlive(StgClosure *p) +{ + StgInfoTable *info; + + while (1) { + + 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 closures in generations that we're not collecting. */ + if (LOOKS_LIKE_STATIC(p) || Bdescr((P_)p)->gen->no > N) { + return p; + } + + switch (info->type) { + + case IND: + case IND_STATIC: + case IND_PERM: + case IND_OLDGEN: /* rely on compatible layout with StgInd */ + case IND_OLDGEN_PERM: + /* follow indirections */ + p = ((StgInd *)p)->indirectee; + continue; + + case EVACUATED: + /* alive! */ + return ((StgEvacuated *)p)->evacuee; + + default: + /* dead. */ + return NULL; + } + } +} + StgClosure * MarkRoot(StgClosure *root) { - root = evacuate(root); - return root; + return evacuate(root); } -static inline void addBlock(step *step) +static void addBlock(step *step) { bdescr *bd = allocBlock(); bd->gen = step->gen; @@ -692,17 +819,16 @@ static inline void addBlock(step *step) } static __inline__ StgClosure * -copy(StgClosure *src, W_ size, bdescr *bd) +copy(StgClosure *src, nat size, step *step) { - step *step; P_ to, from, dest; + TICK_GC_WORDS_COPIED(size); /* Find out where we're going, using the handy "to" pointer in * the step of the source object. If it turns out we need to * evacuate to an older generation, adjust it here (see comment * by evacuate()). */ - step = bd->step->to; if (step->gen->no < evac_gen) { step = &generations[evac_gen].steps[0]; } @@ -714,11 +840,40 @@ copy(StgClosure *src, W_ size, bdescr *bd) addBlock(step); } + for(to = step->hp, from = (P_)src; size>0; --size) { + *to++ = *from++; + } + dest = step->hp; - step->hp += size; - for(to = dest, from = (P_)src; size>0; --size) { + step->hp = to; + return (StgClosure *)dest; +} + +/* Special version of copy() for when we only want to copy the info + * pointer of an object, but reserve some padding after it. This is + * used to optimise evacuation of BLACKHOLEs. + */ + +static __inline__ StgClosure * +copyPart(StgClosure *src, nat size_to_reserve, nat size_to_copy, step *step) +{ + P_ dest, to, from; + + TICK_GC_WORDS_COPIED(size_to_copy); + if (step->gen->no < evac_gen) { + step = &generations[evac_gen].steps[0]; + } + + if (step->hp + size_to_reserve >= step->hpLim) { + addBlock(step); + } + + for(to = step->hp, from = (P_)src; size_to_copy>0; --size_to_copy) { *to++ = *from++; } + + dest = step->hp; + step->hp += size_to_reserve; return (StgClosure *)dest; } @@ -732,25 +887,6 @@ upd_evacuee(StgClosure *p, StgClosure *dest) } /* ----------------------------------------------------------------------------- - Evacuate a mutable object - - If we evacuate a mutable object to an old generation, cons the - object onto the older generation's mutable list. - -------------------------------------------------------------------------- */ - -static inline void -evacuate_mutable(StgMutClosure *c) -{ - bdescr *bd; - - bd = Bdescr((P_)c); - if (bd->gen->no > 0) { - c->mut_link = bd->gen->mut_list; - bd->gen->mut_list = c; - } -} - -/* ----------------------------------------------------------------------------- Evacuate a large object This just consists of removing the object from the (doubly-linked) @@ -777,6 +913,7 @@ evacuate_large(StgPtr p, rtsBool mutable) */ if (bd->gen->no < evac_gen) { failed_to_evac = rtsTrue; + TICK_GC_FAILED_PROMOTION(); } return; } @@ -806,7 +943,7 @@ evacuate_large(StgPtr p, rtsBool mutable) bd->evacuated = 1; if (mutable) { - evacuate_mutable((StgMutClosure *)p); + recordMutable((StgMutClosure *)p); } } @@ -838,7 +975,7 @@ mkMutCons(StgClosure *ptr, generation *gen) SET_HDR(q,&MUT_CONS_info,CCS_GC); q->var = ptr; - evacuate_mutable((StgMutClosure *)q); + recordOldToNewPtrs((StgMutClosure *)q); return (StgClosure *)q; } @@ -874,6 +1011,7 @@ evacuate(StgClosure *q) { StgClosure *to; bdescr *bd = NULL; + step *step; const StgInfoTable *info; loop: @@ -887,9 +1025,11 @@ loop: if (bd->gen->no < evac_gen) { /* nope */ failed_to_evac = rtsTrue; + TICK_GC_FAILED_PROMOTION(); } return q; } + step = bd->step->to; } /* make sure the info pointer is into text space */ @@ -900,15 +1040,45 @@ loop: switch (info -> type) { case BCO: - to = copy(q,bco_sizeW(stgCast(StgBCO*,q)),bd); + to = copy(q,bco_sizeW(stgCast(StgBCO*,q)),step); upd_evacuee(q,to); return to; case MUT_VAR: + ASSERT(q->header.info != &MUT_CONS_info); case MVAR: - to = copy(q,sizeW_fromITBL(info),bd); + to = copy(q,sizeW_fromITBL(info),step); + upd_evacuee(q,to); + recordMutable((StgMutClosure *)to); + return to; + + case STABLE_NAME: + stable_ptr_table[((StgStableName *)q)->sn].keep = rtsTrue; + to = copy(q,sizeofW(StgStableName),step); + upd_evacuee(q,to); + return to; + + case FUN_1_0: + case FUN_0_1: + case CONSTR_1_0: + case CONSTR_0_1: + to = copy(q,sizeofW(StgHeader)+1,step); + upd_evacuee(q,to); + return to; + + case THUNK_1_0: /* here because of MIN_UPD_SIZE */ + case THUNK_0_1: + case FUN_1_1: + case FUN_0_2: + case FUN_2_0: + case THUNK_1_1: + case THUNK_0_2: + case THUNK_2_0: + case CONSTR_1_1: + case CONSTR_0_2: + case CONSTR_2_0: + to = copy(q,sizeofW(StgHeader)+2,step); upd_evacuee(q,to); - evacuate_mutable((StgMutClosure *)to); return to; case FUN: @@ -920,16 +1090,22 @@ loop: case CAF_ENTERED: case WEAK: case FOREIGN: - to = copy(q,sizeW_fromITBL(info),bd); + to = copy(q,sizeW_fromITBL(info),step); upd_evacuee(q,to); return to; case CAF_BLACKHOLE: case BLACKHOLE: - to = copy(q,BLACKHOLE_sizeW(),bd); + to = copyPart(q,BLACKHOLE_sizeW(),sizeofW(StgHeader),step); upd_evacuee(q,to); return to; + case BLACKHOLE_BQ: + to = copy(q,BLACKHOLE_sizeW(),step); + upd_evacuee(q,to); + recordMutable((StgMutClosure *)to); + return to; + case THUNK_SELECTOR: { const StgInfoTable* selectee_info; @@ -939,6 +1115,11 @@ 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: { StgNat32 offset = info->layout.selector_offset; @@ -960,6 +1141,7 @@ loop: if (bd->evacuated) { if (bd->gen->no < evac_gen) { failed_to_evac = rtsTrue; + TICK_GC_FAILED_PROMOTION(); } return q; } @@ -988,12 +1170,18 @@ loop: goto selector_loop; 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 THUNK_SELECTOR: /* aargh - do recursively???? */ case CAF_UNENTERED: case CAF_BLACKHOLE: case BLACKHOLE: + case BLACKHOLE_BQ: /* not evaluated yet */ break; @@ -1001,7 +1189,7 @@ loop: barf("evacuate: THUNK_SELECTOR: strange selectee"); } } - to = copy(q,THUNK_SELECTOR_sizeW(),bd); + to = copy(q,THUNK_SELECTOR_sizeW(),step); upd_evacuee(q,to); return to; @@ -1061,7 +1249,7 @@ loop: case PAP: /* these are special - the payload is a copy of a chunk of stack, tagging and all. */ - to = copy(q,pap_sizeW(stgCast(StgPAP*,q)),bd); + to = copy(q,pap_sizeW(stgCast(StgPAP*,q)),step); upd_evacuee(q,to); return to; @@ -1078,7 +1266,8 @@ loop: if (Bdescr((P_)p)->gen->no < evac_gen) { /* fprintf(stderr,"evac failed!\n");*/ failed_to_evac = rtsTrue; - } + TICK_GC_FAILED_PROMOTION(); + } } return ((StgEvacuated*)q)->evacuee; @@ -1092,7 +1281,7 @@ loop: return q; } else { /* just copy the block */ - to = copy(q,size,bd); + to = copy(q,size,step); upd_evacuee(q,to); return to; } @@ -1108,10 +1297,10 @@ loop: to = q; } else { /* just copy the block */ - to = copy(q,size,bd); + to = copy(q,size,step); upd_evacuee(q,to); if (info->type == MUT_ARR_PTRS) { - evacuate_mutable((StgMutClosure *)to); + recordMutable((StgMutClosure *)to); } } return to; @@ -1126,15 +1315,14 @@ loop: /* Large TSOs don't get moved, so no relocation is required. */ if (size >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) { - evacuate_large((P_)q, rtsFalse); - tso->mut_link = NULL; /* see below */ + evacuate_large((P_)q, rtsTrue); return q; /* To evacuate a small TSO, we need to relocate the update frame * list it contains. */ } else { - StgTSO *new_tso = (StgTSO *)copy((StgClosure *)tso,tso_sizeW(tso),bd); + StgTSO *new_tso = (StgTSO *)copy((StgClosure *)tso,tso_sizeW(tso),step); diff = (StgPtr)new_tso - (StgPtr)tso; /* In *words* */ @@ -1146,14 +1334,7 @@ loop: relocate_TSO(tso, new_tso); upd_evacuee(q,(StgClosure *)new_tso); - /* don't evac_mutable - these things are marked mutable as - * required. We *do* need to zero the mut_link field, though: - * this TSO might have been on the mutable list for this - * generation, but we're collecting this generation anyway so - * we didn't follow the mutable list. - */ - new_tso->mut_link = NULL; - + recordMutable((StgMutClosure *)new_tso); return (StgClosure *)new_tso; } } @@ -1312,6 +1493,54 @@ scavenge(step *step) break; } + case THUNK_2_0: + case FUN_2_0: + scavenge_srt(info); + case CONSTR_2_0: + ((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_1_0: + scavenge_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); + 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); + p += sizeofW(StgHeader) + 2; /* MIN_UPD_SIZE */ + break; + + case FUN_0_1: + scavenge_srt(info); + case CONSTR_0_1: + p += sizeofW(StgHeader) + 1; + break; + + case THUNK_0_2: + case FUN_0_2: + scavenge_srt(info); + case CONSTR_0_2: + p += sizeofW(StgHeader) + 2; + break; + + case THUNK_1_1: + case FUN_1_1: + scavenge_srt(info); + case CONSTR_1_1: + ((StgClosure *)p)->payload[0] = evacuate(((StgClosure *)p)->payload[0]); + p += sizeofW(StgHeader) + 2; + break; + case FUN: case THUNK: scavenge_srt(info); @@ -1320,6 +1549,7 @@ scavenge(step *step) case CONSTR: case WEAK: case FOREIGN: + case STABLE_NAME: case IND_PERM: case IND_OLDGEN_PERM: case CAF_UNENTERED: @@ -1347,10 +1577,18 @@ scavenge(step *step) case CAF_BLACKHOLE: case BLACKHOLE: + p += BLACKHOLE_sizeW(); + break; + + case BLACKHOLE_BQ: { - StgBlackHole *bh = (StgBlackHole *)p; + StgBlockingQueue *bh = (StgBlockingQueue *)p; (StgClosure *)bh->blocking_queue = evacuate((StgClosure *)bh->blocking_queue); + if (failed_to_evac) { + failed_to_evac = rtsFalse; + recordMutable((StgMutClosure *)bh); + } p += BLACKHOLE_sizeW(); break; } @@ -1435,7 +1673,7 @@ scavenge(step *step) } if (failed_to_evac) { /* we can do this easier... */ - evacuate_mutable((StgMutClosure *)start); + recordMutable((StgMutClosure *)start); failed_to_evac = rtsFalse; } break; @@ -1487,21 +1725,36 @@ scavenge(step *step) objects can have this property. -------------------------------------------------------------------------- */ static rtsBool -scavenge_one(StgPtr p) +scavenge_one(StgClosure *p) { StgInfoTable *info; rtsBool no_luck; - ASSERT(p && (LOOKS_LIKE_GHC_INFO(GET_INFO((StgClosure *)p)) - || IS_HUGS_CONSTR_INFO(GET_INFO((StgClosure *)p)))); + ASSERT(p && (LOOKS_LIKE_GHC_INFO(GET_INFO(p)) + || IS_HUGS_CONSTR_INFO(GET_INFO(p)))); - info = get_itbl((StgClosure *)p); + info = get_itbl(p); switch (info -> type) { case FUN: + case FUN_1_0: /* hardly worth specialising these guys */ + case FUN_0_1: + case FUN_1_1: + case FUN_0_2: + case FUN_2_0: case THUNK: + case THUNK_1_0: + case THUNK_0_1: + case THUNK_1_1: + case THUNK_0_2: + case THUNK_2_0: case CONSTR: + case CONSTR_1_0: + case CONSTR_0_1: + case CONSTR_1_1: + case CONSTR_0_2: + case CONSTR_2_0: case WEAK: case FOREIGN: case IND_PERM: @@ -1509,29 +1762,24 @@ scavenge_one(StgPtr p) case CAF_UNENTERED: case CAF_ENTERED: { - StgPtr end; + StgPtr q, end; - end = (P_)((StgClosure *)p)->payload + info->layout.payload.ptrs; - for (p = (P_)((StgClosure *)p)->payload; p < end; p++) { - (StgClosure *)*p = evacuate((StgClosure *)*p); + end = (P_)p->payload + info->layout.payload.ptrs; + for (q = (P_)p->payload; q < end; q++) { + (StgClosure *)*q = evacuate((StgClosure *)*q); } break; } case CAF_BLACKHOLE: case BLACKHOLE: - { - StgBlackHole *bh = (StgBlackHole *)p; - (StgClosure *)bh->blocking_queue = - evacuate((StgClosure *)bh->blocking_queue); break; - } case THUNK_SELECTOR: { StgSelector *s = (StgSelector *)p; s->selectee = evacuate(s->selectee); - break; + break; } case AP_UPD: /* same as PAPs */ @@ -1540,7 +1788,7 @@ scavenge_one(StgPtr p) * evacuate the function pointer too... */ { - StgPAP* pap = stgCast(StgPAP*,p); + StgPAP* pap = (StgPAP *)p; pap->fun = evacuate(pap->fun); scavenge_stack((P_)pap->payload, (P_)pap->payload + pap->n_args); @@ -1574,6 +1822,83 @@ scavenge_one(StgPtr p) -------------------------------------------------------------------------- */ static StgMutClosure * +scavenge_mut_once_list(StgMutClosure *p, nat gen) +{ + StgInfoTable *info; + StgMutClosure *start; + StgMutClosure **prev; + + prev = &start; + start = p; + + evac_gen = gen; + failed_to_evac = rtsFalse; + + for (; p != END_MUT_LIST; p = *prev) { + + /* 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)))); + + info = get_itbl(p); + switch(info->type) { + + case IND_OLDGEN: + case IND_OLDGEN_PERM: + case IND_STATIC: + /* Try to pull the indirectee into this generation, so we can + * remove the indirection from the mutable list. + */ + ((StgIndOldGen *)p)->indirectee = + evacuate(((StgIndOldGen *)p)->indirectee); + + /* failed_to_evac might happen if we've got more than two + * generations, we're collecting only generation 0, the + * indirection resides in generation 2 and the indirectee is + * in generation 1. + */ + if (failed_to_evac) { + failed_to_evac = rtsFalse; + prev = &p->mut_link; + } else { + *prev = p->mut_link; + /* the mut_link field of an IND_STATIC is overloaded as the + * static link field too (it just so happens that we don't need + * both at the same time), so we need to NULL it out when + * removing this object from the mutable list because the static + * link fields are all assumed to be NULL before doing a major + * collection. + */ + p->mut_link = NULL; + } + continue; + + case MUT_VAR: + /* MUT_CONS is a kind of MUT_VAR, except it that we try to remove + * it from the mutable list if possible by promoting whatever it + * points to. + */ + ASSERT(p->header.info == &MUT_CONS_info); + if (scavenge_one(((StgMutVar *)p)->var) == rtsTrue) { + /* didn't manage to promote everything, so leave the + * MUT_CONS on the list. + */ + prev = &p->mut_link; + } else { + *prev = p->mut_link; + } + continue; + + default: + /* shouldn't have anything else on the mutables list */ + barf("scavenge_mut_once_list: strange object?"); + } + } + return start; +} + + +static StgMutClosure * scavenge_mutable_list(StgMutClosure *p, nat gen) { StgInfoTable *info; @@ -1637,23 +1962,21 @@ scavenge_mutable_list(StgMutClosure *p, nat gen) * it from the mutable list if possible by promoting whatever it * points to. */ - if (p->header.info == &MUT_CONS_info) { - evac_gen = gen; - if (scavenge_one((P_)((StgMutVar *)p)->var) == rtsTrue) { - /* didn't manage to promote everything, so leave the - * MUT_CONS on the list. - */ - prev = &p->mut_link; - } else { - *prev = p->mut_link; - } - evac_gen = 0; - } else { - ((StgMutVar *)p)->var = evacuate(((StgMutVar *)p)->var); - prev = &p->mut_link; - } + ASSERT(p->header.info != &MUT_CONS_info); + ((StgMutVar *)p)->var = evacuate(((StgMutVar *)p)->var); + prev = &p->mut_link; continue; + case MVAR: + { + StgMVar *mvar = (StgMVar *)p; + (StgClosure *)mvar->head = evacuate((StgClosure *)mvar->head); + (StgClosure *)mvar->tail = evacuate((StgClosure *)mvar->tail); + (StgClosure *)mvar->value = evacuate((StgClosure *)mvar->value); + prev = &p->mut_link; + continue; + } + case TSO: /* follow ptrs and remove this from the mutable list */ { @@ -1677,36 +2000,18 @@ scavenge_mutable_list(StgMutClosure *p, nat gen) continue; } - case IND_OLDGEN: - case IND_OLDGEN_PERM: - case IND_STATIC: - /* Try to pull the indirectee into this generation, so we can - * remove the indirection from the mutable list. - */ - evac_gen = gen; - ((StgIndOldGen *)p)->indirectee = - evacuate(((StgIndOldGen *)p)->indirectee); - evac_gen = 0; - - if (failed_to_evac) { - failed_to_evac = rtsFalse; + case BLACKHOLE_BQ: + { + StgBlockingQueue *bh = (StgBlockingQueue *)p; + (StgClosure *)bh->blocking_queue = + evacuate((StgClosure *)bh->blocking_queue); prev = &p->mut_link; - } else { - *prev = p->mut_link; - /* the mut_link field of an IND_STATIC is overloaded as the - * static link field too (it just so happens that we don't need - * both at the same time), so we need to NULL it out when - * removing this object from the mutable list because the static - * link fields are all assumed to be NULL before doing a major - * collection. - */ - p->mut_link = NULL; + break; } - continue; - + default: /* shouldn't have anything else on the mutables list */ - barf("scavenge_mutable_object: non-mutable object?"); + barf("scavenge_mut_list: strange object?"); } } return start; @@ -1745,6 +2050,18 @@ scavenge_static(void) { StgInd *ind = (StgInd *)p; ind->indirectee = evacuate(ind->indirectee); + + /* might fail to evacuate it, in which case we have to pop it + * back on the mutable list (and take it off the + * scavenged_static list because the static link and mut link + * pointers are one and the same). + */ + if (failed_to_evac) { + failed_to_evac = rtsFalse; + scavenged_static_objects = STATIC_LINK(info,p); + ((StgMutClosure *)ind)->mut_link = oldest_gen->mut_once_list; + oldest_gen->mut_once_list = (StgMutClosure *)ind; + } break; } @@ -1769,6 +2086,8 @@ scavenge_static(void) barf("scavenge_static"); } + ASSERT(failed_to_evac == rtsFalse); + /* get the next static object from the list. Remeber, there might * be more stuff on this list now that we've done some evacuating! * (static_objects is a global) @@ -1844,8 +2163,8 @@ scavenge_stack(StgPtr p, StgPtr stack_end) goto follow_srt; /* Specialised code for update frames, since they're so common. - * We *know* the updatee points to a BLACKHOLE or CAF_BLACKHOLE, - * so just inline the code to evacuate it here. + * 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: { @@ -1859,12 +2178,31 @@ scavenge_stack(StgPtr p, StgPtr stack_end) continue; } else { bdescr *bd = Bdescr((P_)frame->updatee); - ASSERT(type == BLACKHOLE || type == CAF_BLACKHOLE); - if (bd->gen->no >= evac_gen && bd->gen->no > N) { continue; } - to = copy(frame->updatee, BLACKHOLE_sizeW(), bd); - upd_evacuee(frame->updatee,to); - frame->updatee = to; - continue; + step *step; + if (bd->gen->no > N) { + if (bd->gen->no < evac_gen) { + failed_to_evac = rtsTrue; + } + continue; + } + step = bd->step->to; + switch (type) { + case BLACKHOLE: + case CAF_BLACKHOLE: + to = copyPart(frame->updatee, BLACKHOLE_sizeW(), + sizeofW(StgHeader), step); + upd_evacuee(frame->updatee,to); + frame->updatee = to; + continue; + case BLACKHOLE_BQ: + to = copy(frame->updatee, BLACKHOLE_sizeW(), step); + upd_evacuee(frame->updatee,to); + frame->updatee = to; + recordMutable((StgMutClosure *)to); + continue; + default: + barf("scavenge_stack: UPDATE_FRAME updatee"); + } } } @@ -1995,7 +2333,7 @@ scavenge_large(step *step) } evac_gen = 0; if (failed_to_evac) { - evacuate_mutable((StgMutClosure *)start); + recordMutable((StgMutClosure *)start); } continue; } @@ -2172,7 +2510,7 @@ static void threadLazyBlackHole(StgTSO *tso) { StgUpdateFrame *update_frame; - StgBlackHole *bh; + StgBlockingQueue *bh; StgPtr stack_end; stack_end = &tso->stack[tso->stack_size]; @@ -2186,21 +2524,22 @@ threadLazyBlackHole(StgTSO *tso) break; case UPDATE_FRAME: - bh = stgCast(StgBlackHole*,update_frame->updatee); + 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 == &BLACKHOLE_info) { + return; + } - /* Don't for now: when we enter a CAF, we create a black hole on - * the heap and make the update frame point to it. Thus the - * above optimisation doesn't apply. - */ - if (bh->header.info != &BLACKHOLE_info - && bh->header.info != &CAF_BLACKHOLE_info) { + if (bh->header.info != &BLACKHOLE_BQ_info && + bh->header.info != &CAF_BLACKHOLE_info) { SET_INFO(bh,&BLACKHOLE_info); - bh->blocking_queue = END_TSO_QUEUE; } update_frame = update_frame->link; @@ -2249,8 +2588,8 @@ threadSqueezeStack(StgTSO *tso) * added to the stack, rather than the way we see them in this * walk. (It makes the next loop less confusing.) * - * Could stop if we find an update frame pointing to a black hole, - * but see comment in threadLazyBlackHole(). + * Stop if we find an update frame pointing to a black hole + * (see comment in threadLazyBlackHole()). */ next_frame = NULL; @@ -2259,6 +2598,10 @@ threadSqueezeStack(StgTSO *tso) frame->link = next_frame; next_frame = frame; frame = prev_frame; + if (get_itbl(frame)->type == UPDATE_FRAME + && frame->updatee->header.info == &BLACKHOLE_info) { + break; + } } /* Now, we're at the bottom. Frame points to the lowest update @@ -2318,20 +2661,16 @@ threadSqueezeStack(StgTSO *tso) * slower --SDM */ #if 0 /* do it properly... */ - if (GET_INFO(updatee_bypass) == BLACKHOLE_info - || GET_INFO(updatee_bypass) == CAF_BLACKHOLE_info - ) { + if (GET_INFO(updatee_bypass) == BLACKHOLE_BQ_info) { /* Sigh. It has one. Don't lose those threads! */ - if (GET_INFO(updatee_keep) == BLACKHOLE_info - || GET_INFO(updatee_keep) == CAF_BLACKHOLE_info - ) { + if (GET_INFO(updatee_keep) == BLACKHOLE_BQ_info) { /* Urgh. Two queues. Merge them. */ - P_ keep_tso = ((StgBlackHole *)updatee_keep)->blocking_queue; + P_ keep_tso = ((StgBlockingQueue *)updatee_keep)->blocking_queue; while (keep_tso->link != END_TSO_QUEUE) { keep_tso = keep_tso->link; } - keep_tso->link = ((StgBlackHole *)updatee_bypass)->blocking_queue; + keep_tso->link = ((StgBlockingQueue *)updatee_bypass)->blocking_queue; } else { /* For simplicity, just swap the BQ for the BH */ @@ -2359,12 +2698,10 @@ threadSqueezeStack(StgTSO *tso) /* Do lazy black-holing. */ if (is_update_frame) { - StgBlackHole *bh = (StgBlackHole *)frame->updatee; - if (bh->header.info != &BLACKHOLE_info - && bh->header.info != &CAF_BLACKHOLE_info - ) { + StgBlockingQueue *bh = (StgBlockingQueue *)frame->updatee; + if (bh->header.info != &BLACKHOLE_BQ_info && + bh->header.info != &CAF_BLACKHOLE_info) { SET_INFO(bh,&BLACKHOLE_info); - bh->blocking_queue = END_TSO_QUEUE; } }