From d13df738cbbe8017ae19ae2702f4e10805ee521b Mon Sep 17 00:00:00 2001 From: Simon Marlow Date: Wed, 31 Oct 2007 15:33:39 +0000 Subject: [PATCH] compile special minor GC versions of evacuate() and scavenge_block() This is for two reasons: minor GCs don't need to do per-object locking for parallel GC, which is fairly expensive, and secondly minor GCs don't need to follow SRTs. --- rts/sm/Evac.c | 460 +++---------------------------------------- rts/sm/Evac.c-inc | 565 +++++++++++++++++++++++++++++++++++++++++++++++++++++ rts/sm/Evac.h | 3 +- rts/sm/Scav.c | 446 ++---------------------------------------- rts/sm/Scav.c-inc | 472 ++++++++++++++++++++++++++++++++++++++++++++ 5 files changed, 1080 insertions(+), 866 deletions(-) create mode 100644 rts/sm/Evac.c-inc create mode 100644 rts/sm/Scav.c-inc diff --git a/rts/sm/Evac.c b/rts/sm/Evac.c index 7468b4e..e8f85b6 100644 --- a/rts/sm/Evac.c +++ b/rts/sm/Evac.c @@ -26,6 +26,11 @@ #define MAX_THUNK_SELECTOR_DEPTH 16 static void eval_thunk_selector (StgClosure **q, StgSelector * p, rtsBool); +STATIC_INLINE void evacuate_large(StgPtr p); + +/* ----------------------------------------------------------------------------- + Allocate some space in which to copy an object. + -------------------------------------------------------------------------- */ STATIC_INLINE StgPtr alloc_for_copy (nat size, step *stp) @@ -176,6 +181,16 @@ copy(StgClosure **p, StgClosure *src, nat size, step *stp) } /* ----------------------------------------------------------------------------- + The evacuate() code + -------------------------------------------------------------------------- */ + +#define MINOR_GC +#include "Evac.c-inc" + +#undef MINOR_GC +#include "Evac.c-inc" + +/* ----------------------------------------------------------------------------- Evacuate a large object This just consists of removing the object from the (doubly-linked) @@ -186,7 +201,6 @@ copy(StgClosure **p, StgClosure *src, nat size, step *stp) that has been evacuated, or unset otherwise. -------------------------------------------------------------------------- */ - STATIC_INLINE void evacuate_large(StgPtr p) { @@ -244,434 +258,16 @@ evacuate_large(StgPtr p) } /* ----------------------------------------------------------------------------- - Evacuate - - This is called (eventually) for every live object in the system. - - The caller to evacuate specifies a desired generation in the - gct->evac_step thread-local variable. The following conditions apply to - evacuating an object which resides in generation M when we're - collecting up to generation N - - if M >= gct->evac_step - if M > N do nothing - else evac to step->to - - if M < gct->evac_step evac to gct->evac_step, step 0 - - if the object is already evacuated, then we check which generation - it now resides in. - - if M >= gct->evac_step do nothing - if M < gct->evac_step set gct->failed_to_evac flag to indicate that we - didn't manage to evacuate this object into gct->evac_step. - - - OPTIMISATION NOTES: + Evaluate a THUNK_SELECTOR if possible. - evacuate() is the single most important function performance-wise - in the GC. Various things have been tried to speed it up, but as - far as I can tell the code generated by gcc 3.2 with -O2 is about - as good as it's going to get. We pass the argument to evacuate() - in a register using the 'regparm' attribute (see the prototype for - evacuate() near the top of this file). + p points to a THUNK_SELECTOR that we want to evaluate. The + result of "evaluating" it will be evacuated and a pointer to the + to-space closure will be returned. - Changing evacuate() to take an (StgClosure **) rather than - returning the new pointer seems attractive, because we can avoid - writing back the pointer when it hasn't changed (eg. for a static - object, or an object in a generation > N). However, I tried it and - it doesn't help. One reason is that the (StgClosure **) pointer - gets spilled to the stack inside evacuate(), resulting in far more - extra reads/writes than we save. + If the THUNK_SELECTOR could not be evaluated (its selectee is still + a THUNK, for example), then the THUNK_SELECTOR itself will be + evacuated. -------------------------------------------------------------------------- */ - -REGPARM1 void -evacuate(StgClosure **p) -{ - bdescr *bd = NULL; - step *stp; - StgClosure *q; - const StgInfoTable *info; - StgWord tag; - - q = *p; - -loop: - /* The tag and the pointer are split, to be merged after evacing */ - tag = GET_CLOSURE_TAG(q); - q = UNTAG_CLOSURE(q); - - ASSERT(LOOKS_LIKE_CLOSURE_PTR(q)); - - if (!HEAP_ALLOCED(q)) { - - if (!major_gc) return; - - info = get_itbl(q); - switch (info->type) { - - case THUNK_STATIC: - if (info->srt_bitmap != 0 && - *THUNK_STATIC_LINK((StgClosure *)q) == NULL) { - ACQUIRE_SPIN_LOCK(&static_objects_sync); - if (*THUNK_STATIC_LINK((StgClosure *)q) == NULL) { - *THUNK_STATIC_LINK((StgClosure *)q) = static_objects; - static_objects = (StgClosure *)q; - } - RELEASE_SPIN_LOCK(&static_objects_sync); - } - return; - - case FUN_STATIC: - if (info->srt_bitmap != 0 && - *FUN_STATIC_LINK((StgClosure *)q) == NULL) { - ACQUIRE_SPIN_LOCK(&static_objects_sync); - if (*FUN_STATIC_LINK((StgClosure *)q) == NULL) { - *FUN_STATIC_LINK((StgClosure *)q) = static_objects; - static_objects = (StgClosure *)q; - } - RELEASE_SPIN_LOCK(&static_objects_sync); - } - return; - - case IND_STATIC: - /* If q->saved_info != NULL, then it's a revertible CAF - it'll be - * on the CAF list, so don't do anything with it here (we'll - * scavenge it later). - */ - if (((StgIndStatic *)q)->saved_info == NULL) { - ACQUIRE_SPIN_LOCK(&static_objects_sync); - if (*IND_STATIC_LINK((StgClosure *)q) == NULL) { - *IND_STATIC_LINK((StgClosure *)q) = static_objects; - static_objects = (StgClosure *)q; - } - RELEASE_SPIN_LOCK(&static_objects_sync); - } - return; - - case CONSTR_STATIC: - if (*STATIC_LINK(info,(StgClosure *)q) == NULL) { - ACQUIRE_SPIN_LOCK(&static_objects_sync); - // re-test, after acquiring lock - if (*STATIC_LINK(info,(StgClosure *)q) == NULL) { - *STATIC_LINK(info,(StgClosure *)q) = static_objects; - static_objects = (StgClosure *)q; - } - RELEASE_SPIN_LOCK(&static_objects_sync); - /* I am assuming that static_objects pointers are not - * written to other objects, and thus, no need to retag. */ - } - return; - - case CONSTR_NOCAF_STATIC: - /* no need to put these on the static linked list, they don't need - * to be scavenged. - */ - return; - - default: - barf("evacuate(static): strange closure type %d", (int)(info->type)); - } - } - - bd = Bdescr((P_)q); - - if (bd->gen_no > N) { - /* Can't evacuate this object, because it's in a generation - * older than the ones we're collecting. Let's hope that it's - * in gct->evac_step or older, or we will have to arrange to track - * this pointer using the mutable list. - */ - if (bd->step < gct->evac_step) { - // nope - gct->failed_to_evac = rtsTrue; - TICK_GC_FAILED_PROMOTION(); - } - return; - } - - if ((bd->flags & (BF_LARGE | BF_COMPACTED | BF_EVACUATED)) != 0) { - - /* pointer into to-space: just return it. This normally - * shouldn't happen, but alllowing it makes certain things - * slightly easier (eg. the mutable list can contain the same - * object twice, for example). - */ - if (bd->flags & BF_EVACUATED) { - if (bd->step < gct->evac_step) { - gct->failed_to_evac = rtsTrue; - TICK_GC_FAILED_PROMOTION(); - } - return; - } - - /* evacuate large objects by re-linking them onto a different list. - */ - if (bd->flags & BF_LARGE) { - info = get_itbl(q); - if (info->type == TSO && - ((StgTSO *)q)->what_next == ThreadRelocated) { - q = (StgClosure *)((StgTSO *)q)->link; - *p = q; - goto loop; - } - evacuate_large((P_)q); - return; - } - - /* If the object is in a step that we're compacting, then we - * need to use an alternative evacuate procedure. - */ - if (bd->flags & BF_COMPACTED) { - if (!is_marked((P_)q,bd)) { - mark((P_)q,bd); - if (mark_stack_full()) { - mark_stack_overflowed = rtsTrue; - reset_mark_stack(); - } - push_mark_stack((P_)q); - } - return; - } - } - - stp = bd->step->to; - - info = get_itbl(q); - - switch (info->type) { - - case WHITEHOLE: - goto loop; - - case MUT_VAR_CLEAN: - case MUT_VAR_DIRTY: - case MVAR_CLEAN: - case MVAR_DIRTY: - copy(p,q,sizeW_fromITBL(info),stp); - return; - - case CONSTR_0_1: - { - StgWord w = (StgWord)q->payload[0]; - if (q->header.info == Czh_con_info && - // unsigned, so always true: (StgChar)w >= MIN_CHARLIKE && - (StgChar)w <= MAX_CHARLIKE) { - *p = TAG_CLOSURE(tag, - (StgClosure *)CHARLIKE_CLOSURE((StgChar)w) - ); - } - if (q->header.info == Izh_con_info && - (StgInt)w >= MIN_INTLIKE && (StgInt)w <= MAX_INTLIKE) { - *p = TAG_CLOSURE(tag, - (StgClosure *)INTLIKE_CLOSURE((StgInt)w) - ); - } - else { - copy_tag(p,q,sizeofW(StgHeader)+1,stp,tag); - } - return; - } - - case FUN_0_1: - case FUN_1_0: - case CONSTR_1_0: - copy_tag(p,q,sizeofW(StgHeader)+1,stp,tag); - return; - - case THUNK_1_0: - case THUNK_0_1: - copy(p,q,sizeofW(StgThunk)+1,stp); - return; - - case THUNK_1_1: - case THUNK_2_0: - case THUNK_0_2: -#ifdef NO_PROMOTE_THUNKS - if (bd->gen_no == 0 && - bd->step->no != 0 && - bd->step->no == generations[bd->gen_no].n_steps-1) { - stp = bd->step; - } -#endif - copy(p,q,sizeofW(StgThunk)+2,stp); - return; - - case FUN_1_1: - case FUN_2_0: - case FUN_0_2: - case CONSTR_1_1: - case CONSTR_2_0: - copy_tag(p,q,sizeofW(StgHeader)+2,stp,tag); - return; - - case CONSTR_0_2: - copy_tag(p,q,sizeofW(StgHeader)+2,stp,tag); - return; - - case THUNK: - copy(p,q,thunk_sizeW_fromITBL(info),stp); - return; - - case FUN: - case IND_PERM: - case IND_OLDGEN_PERM: - case WEAK: - case STABLE_NAME: - case CONSTR: - copy_tag(p,q,sizeW_fromITBL(info),stp,tag); - return; - - case BCO: - copy(p,q,bco_sizeW((StgBCO *)q),stp); - return; - - case CAF_BLACKHOLE: - case SE_CAF_BLACKHOLE: - case SE_BLACKHOLE: - case BLACKHOLE: - copyPart(p,q,BLACKHOLE_sizeW(),sizeofW(StgHeader),stp); - return; - - case THUNK_SELECTOR: - eval_thunk_selector(p, (StgSelector *)q, rtsTrue); - return; - - case IND: - case IND_OLDGEN: - // follow chains of indirections, don't evacuate them - q = ((StgInd*)q)->indirectee; - *p = q; - goto loop; - - case RET_BCO: - case RET_SMALL: - case RET_BIG: - case RET_DYN: - case UPDATE_FRAME: - case STOP_FRAME: - case CATCH_FRAME: - case CATCH_STM_FRAME: - case CATCH_RETRY_FRAME: - case ATOMICALLY_FRAME: - // shouldn't see these - barf("evacuate: stack frame at %p\n", q); - - case PAP: - copy(p,q,pap_sizeW((StgPAP*)q),stp); - return; - - case AP: - copy(p,q,ap_sizeW((StgAP*)q),stp); - return; - - case AP_STACK: - copy(p,q,ap_stack_sizeW((StgAP_STACK*)q),stp); - return; - - case EVACUATED: - /* Already evacuated, just return the forwarding address. - * HOWEVER: if the requested destination generation (gct->evac_step) is - * older than the actual generation (because the object was - * already evacuated to a younger generation) then we have to - * set the gct->failed_to_evac flag to indicate that we couldn't - * manage to promote the object to the desired generation. - */ - /* - * Optimisation: the check is fairly expensive, but we can often - * shortcut it if either the required generation is 0, or the - * current object (the EVACUATED) is in a high enough generation. - * We know that an EVACUATED always points to an object in the - * same or an older generation. stp is the lowest step that the - * current object would be evacuated to, so we only do the full - * check if stp is too low. - */ - { - StgClosure *e = ((StgEvacuated*)q)->evacuee; - *p = e; - if (stp < gct->evac_step) { // optimisation - if (HEAP_ALLOCED(e) && Bdescr((P_)e)->step < gct->evac_step) { - gct->failed_to_evac = rtsTrue; - TICK_GC_FAILED_PROMOTION(); - } - } - return; - } - - case ARR_WORDS: - // just copy the block - copy(p,q,arr_words_sizeW((StgArrWords *)q),stp); - return; - - case MUT_ARR_PTRS_CLEAN: - case MUT_ARR_PTRS_DIRTY: - case MUT_ARR_PTRS_FROZEN: - case MUT_ARR_PTRS_FROZEN0: - // just copy the block - copy(p,q,mut_arr_ptrs_sizeW((StgMutArrPtrs *)q),stp); - return; - - case TSO: - { - StgTSO *tso = (StgTSO *)q; - - /* Deal with redirected TSOs (a TSO that's had its stack enlarged). - */ - if (tso->what_next == ThreadRelocated) { - q = (StgClosure *)tso->link; - *p = q; - goto loop; - } - - /* To evacuate a small TSO, we need to relocate the update frame - * list it contains. - */ - { - StgTSO *new_tso; - StgPtr r, s; - - copyPart(p,(StgClosure *)tso, tso_sizeW(tso), sizeofW(StgTSO), stp); - new_tso = (StgTSO *)*p; - move_TSO(tso, new_tso); - for (r = tso->sp, s = new_tso->sp; - r < tso->stack+tso->stack_size;) { - *s++ = *r++; - } - return; - } - } - - case TREC_HEADER: - copy(p,q,sizeofW(StgTRecHeader),stp); - return; - - case TVAR_WATCH_QUEUE: - copy(p,q,sizeofW(StgTVarWatchQueue),stp); - return; - - case TVAR: - copy(p,q,sizeofW(StgTVar),stp); - return; - - case TREC_CHUNK: - copy(p,q,sizeofW(StgTRecChunk),stp); - return; - - case ATOMIC_INVARIANT: - copy(p,q,sizeofW(StgAtomicInvariant),stp); - return; - - case INVARIANT_CHECK_QUEUE: - copy(p,q,sizeofW(StgInvariantCheckQueue),stp); - return; - - default: - barf("evacuate: strange closure type %d", (int)(info->type)); - } - - barf("evacuate"); -} - static void unchain_thunk_selectors(StgSelector *p, StgClosure *val) { @@ -700,18 +296,6 @@ unchain_thunk_selectors(StgSelector *p, StgClosure *val) } } -/* ----------------------------------------------------------------------------- - Evaluate a THUNK_SELECTOR if possible. - - p points to a THUNK_SELECTOR that we want to evaluate. The - result of "evaluating" it will be evacuated and a pointer to the - to-space closure will be returned. - - If the THUNK_SELECTOR could not be evaluated (its selectee is still - a THUNK, for example), then the THUNK_SELECTOR itself will be - evacuated. - -------------------------------------------------------------------------- */ - static void eval_thunk_selector (StgClosure **q, StgSelector * p, rtsBool evac) // NB. for legacy reasons, p & q are swapped around :( diff --git a/rts/sm/Evac.c-inc b/rts/sm/Evac.c-inc new file mode 100644 index 0000000..6fa93d1 --- /dev/null +++ b/rts/sm/Evac.c-inc @@ -0,0 +1,565 @@ +/* ----------------------------------------------------------------------------- + * + * (c) The GHC Team 1998-2006 + * + * Generational garbage collector: evacuation functions + * + * ---------------------------------------------------------------------------*/ + +// We have two versions of evacuate(): one for minor GC, and one for +// non-minor, parallel, GC. This file contains the code for both, +// controllled by the CPP symbol MINOR_GC. + +#ifdef MINOR_GC +#define copy(a,b,c,d) copy0(a,b,c,d) +#define copy_tag(a,b,c,d,e) copy_tag0(a,b,c,d,e) +#define copyPart(a,b,c,d,e) copyPart0(a,b,c,d,e) +#define evacuate(a) evacuate0(a) +#else +#undef copy +#undef copy_tag +#undef copyPart +#undef evacuate +#endif + +STATIC_INLINE void +copy_tag(StgClosure **p, StgClosure *src, nat size, step *stp) +{ + StgPtr to, tagged_to, from; + nat i; + StgWord info; + +#if !defined(MINOR_GC) && defined(THREADED_RTS) + do { + info = xchg((StgPtr)&src->header.info, (W_)&stg_WHITEHOLE_info); + // so.. what is it? + } while (info == (W_)&stg_WHITEHOLE_info); + if (info == (W_)&stg_EVACUATED_info) { + src->header.info = (const StgInfoTable *)info; + return evacuate(p); // does the failed_to_evac stuff + } +#else + info = (W_)src->header.info; + src->header.info = &stg_EVACUATED_info; +#endif + + to = alloc_for_copy(size,stp); + tagged_to = (StgPtr)TAG_CLOSURE(tag,(StgClosure*)to); + *p = (StgClosure *)tagged_to; + + TICK_GC_WORDS_COPIED(size); + + from = (StgPtr)src; + to[0] = info; + for (i = 1; i < size; i++) { // unroll for small i + to[i] = from[i]; + } + +// if (to+size+2 < bd->start + BLOCK_SIZE_W) { +// __builtin_prefetch(to + size + 2, 1); +// } + + ((StgEvacuated*)from)->evacuee = (StgClosure *)tagged_to; +#if !defined(MINOR_GC) && defined(THREADED_RTS) + write_barrier(); + ((StgEvacuated*)from)->header.info = &stg_EVACUATED_info; +#endif + +#ifdef PROFILING + // We store the size of the just evacuated object in the LDV word so that + // the profiler can guess the position of the next object later. + SET_EVACUAEE_FOR_LDV(from, size); +#endif +} + + +/* 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 void +copyPart(StgClosure **p, StgClosure *src, nat size_to_reserve, nat size_to_copy, step *stp) +{ + StgPtr to, from; + nat i; + StgWord info; + +#if !defined(MINOR_GC) && defined(THREADED_RTS) + do { + info = xchg((StgPtr)&src->header.info, (W_)&stg_WHITEHOLE_info); + } while (info == (W_)&stg_WHITEHOLE_info); + if (info == (W_)&stg_EVACUATED_info) { + src->header.info = (const StgInfoTable *)info; + return evacuate(p); // does the failed_to_evac stuff + } +#else + info = (W_)src->header.info; + src->header.info = &stg_EVACUATED_info; +#endif + + to = alloc_for_copy(size_to_reserve, stp); + *p = (StgClosure *)to; + + TICK_GC_WORDS_COPIED(size_to_copy); + + from = (StgPtr)src; + to[0] = info; + for (i = 1; i < size_to_copy; i++) { // unroll for small i + to[i] = from[i]; + } + + ((StgEvacuated*)from)->evacuee = (StgClosure *)to; +#if !defined(MINOR_GC) && defined(THREADED_RTS) + write_barrier(); + ((StgEvacuated*)from)->header.info = &stg_EVACUATED_info; +#endif + +#ifdef PROFILING + // We store the size of the just evacuated object in the LDV word so that + // the profiler can guess the position of the next object later. + SET_EVACUAEE_FOR_LDV(from, size_to_reserve); + // fill the slop + if (size_to_reserve - size_to_copy > 0) + LDV_FILL_SLOP(to + size_to_copy - 1, (int)(size_to_reserve - size_to_copy)); +#endif +} + + +/* Copy wrappers that don't tag the closure after copying */ +STATIC_INLINE void +copy(StgClosure **p, StgClosure *src, nat size, step *stp) +{ + copy_tag(p,src,size,stp,0); +} + +/* ---------------------------------------------------------------------------- + Evacuate + + This is called (eventually) for every live object in the system. + + The caller to evacuate specifies a desired generation in the + gct->evac_step thread-local variable. The following conditions apply to + evacuating an object which resides in generation M when we're + collecting up to generation N + + if M >= gct->evac_step + if M > N do nothing + else evac to step->to + + if M < gct->evac_step evac to gct->evac_step, step 0 + + if the object is already evacuated, then we check which generation + it now resides in. + + if M >= gct->evac_step do nothing + if M < gct->evac_step set gct->failed_to_evac flag to indicate that we + didn't manage to evacuate this object into gct->evac_step. + + + OPTIMISATION NOTES: + + evacuate() is the single most important function performance-wise + in the GC. Various things have been tried to speed it up, but as + far as I can tell the code generated by gcc 3.2 with -O2 is about + as good as it's going to get. We pass the argument to evacuate() + in a register using the 'regparm' attribute (see the prototype for + evacuate() near the top of this file). + + Changing evacuate() to take an (StgClosure **) rather than + returning the new pointer seems attractive, because we can avoid + writing back the pointer when it hasn't changed (eg. for a static + object, or an object in a generation > N). However, I tried it and + it doesn't help. One reason is that the (StgClosure **) pointer + gets spilled to the stack inside evacuate(), resulting in far more + extra reads/writes than we save. + ------------------------------------------------------------------------- */ + +REGPARM1 void +evacuate(StgClosure **p) +{ + bdescr *bd = NULL; + step *stp; + StgClosure *q; + const StgInfoTable *info; + StgWord tag; + + q = *p; + +loop: + /* The tag and the pointer are split, to be merged after evacing */ + tag = GET_CLOSURE_TAG(q); + q = UNTAG_CLOSURE(q); + + ASSERT(LOOKS_LIKE_CLOSURE_PTR(q)); + + if (!HEAP_ALLOCED(q)) { + +#ifdef MINOR_GC + return; +#endif + if (!major_gc) return; + + info = get_itbl(q); + switch (info->type) { + + case THUNK_STATIC: + if (info->srt_bitmap != 0 && + *THUNK_STATIC_LINK((StgClosure *)q) == NULL) { + ACQUIRE_SPIN_LOCK(&static_objects_sync); + if (*THUNK_STATIC_LINK((StgClosure *)q) == NULL) { + *THUNK_STATIC_LINK((StgClosure *)q) = static_objects; + static_objects = (StgClosure *)q; + } + RELEASE_SPIN_LOCK(&static_objects_sync); + } + return; + + case FUN_STATIC: + if (info->srt_bitmap != 0 && + *FUN_STATIC_LINK((StgClosure *)q) == NULL) { + ACQUIRE_SPIN_LOCK(&static_objects_sync); + if (*FUN_STATIC_LINK((StgClosure *)q) == NULL) { + *FUN_STATIC_LINK((StgClosure *)q) = static_objects; + static_objects = (StgClosure *)q; + } + RELEASE_SPIN_LOCK(&static_objects_sync); + } + return; + + case IND_STATIC: + /* If q->saved_info != NULL, then it's a revertible CAF - it'll be + * on the CAF list, so don't do anything with it here (we'll + * scavenge it later). + */ + if (((StgIndStatic *)q)->saved_info == NULL) { + ACQUIRE_SPIN_LOCK(&static_objects_sync); + if (*IND_STATIC_LINK((StgClosure *)q) == NULL) { + *IND_STATIC_LINK((StgClosure *)q) = static_objects; + static_objects = (StgClosure *)q; + } + RELEASE_SPIN_LOCK(&static_objects_sync); + } + return; + + case CONSTR_STATIC: + if (*STATIC_LINK(info,(StgClosure *)q) == NULL) { + ACQUIRE_SPIN_LOCK(&static_objects_sync); + // re-test, after acquiring lock + if (*STATIC_LINK(info,(StgClosure *)q) == NULL) { + *STATIC_LINK(info,(StgClosure *)q) = static_objects; + static_objects = (StgClosure *)q; + } + RELEASE_SPIN_LOCK(&static_objects_sync); + /* I am assuming that static_objects pointers are not + * written to other objects, and thus, no need to retag. */ + } + return; + + case CONSTR_NOCAF_STATIC: + /* no need to put these on the static linked list, they don't need + * to be scavenged. + */ + return; + + default: + barf("evacuate(static): strange closure type %d", (int)(info->type)); + } + } + + bd = Bdescr((P_)q); + + if (bd->gen_no > N) { + /* Can't evacuate this object, because it's in a generation + * older than the ones we're collecting. Let's hope that it's + * in gct->evac_step or older, or we will have to arrange to track + * this pointer using the mutable list. + */ + if (bd->step < gct->evac_step) { + // nope + gct->failed_to_evac = rtsTrue; + TICK_GC_FAILED_PROMOTION(); + } + return; + } + + if ((bd->flags & (BF_LARGE | BF_COMPACTED | BF_EVACUATED)) != 0) { + + /* pointer into to-space: just return it. This normally + * shouldn't happen, but alllowing it makes certain things + * slightly easier (eg. the mutable list can contain the same + * object twice, for example). + */ + if (bd->flags & BF_EVACUATED) { + if (bd->step < gct->evac_step) { + gct->failed_to_evac = rtsTrue; + TICK_GC_FAILED_PROMOTION(); + } + return; + } + + /* evacuate large objects by re-linking them onto a different list. + */ + if (bd->flags & BF_LARGE) { + info = get_itbl(q); + if (info->type == TSO && + ((StgTSO *)q)->what_next == ThreadRelocated) { + q = (StgClosure *)((StgTSO *)q)->link; + *p = q; + goto loop; + } + evacuate_large((P_)q); + return; + } + + /* If the object is in a step that we're compacting, then we + * need to use an alternative evacuate procedure. + */ + if (bd->flags & BF_COMPACTED) { + if (!is_marked((P_)q,bd)) { + mark((P_)q,bd); + if (mark_stack_full()) { + mark_stack_overflowed = rtsTrue; + reset_mark_stack(); + } + push_mark_stack((P_)q); + } + return; + } + } + + stp = bd->step->to; + + info = get_itbl(q); + + switch (info->type) { + + case WHITEHOLE: + goto loop; + + case MUT_VAR_CLEAN: + case MUT_VAR_DIRTY: + case MVAR_CLEAN: + case MVAR_DIRTY: + copy(p,q,sizeW_fromITBL(info),stp); + return; + + case CONSTR_0_1: + { + StgWord w = (StgWord)q->payload[0]; + if (q->header.info == Czh_con_info && + // unsigned, so always true: (StgChar)w >= MIN_CHARLIKE && + (StgChar)w <= MAX_CHARLIKE) { + *p = TAG_CLOSURE(tag, + (StgClosure *)CHARLIKE_CLOSURE((StgChar)w) + ); + } + if (q->header.info == Izh_con_info && + (StgInt)w >= MIN_INTLIKE && (StgInt)w <= MAX_INTLIKE) { + *p = TAG_CLOSURE(tag, + (StgClosure *)INTLIKE_CLOSURE((StgInt)w) + ); + } + else { + copy_tag(p,q,sizeofW(StgHeader)+1,stp,tag); + } + return; + } + + case FUN_0_1: + case FUN_1_0: + case CONSTR_1_0: + copy_tag(p,q,sizeofW(StgHeader)+1,stp,tag); + return; + + case THUNK_1_0: + case THUNK_0_1: + copy(p,q,sizeofW(StgThunk)+1,stp); + return; + + case THUNK_1_1: + case THUNK_2_0: + case THUNK_0_2: +#ifdef NO_PROMOTE_THUNKS + if (bd->gen_no == 0 && + bd->step->no != 0 && + bd->step->no == generations[bd->gen_no].n_steps-1) { + stp = bd->step; + } +#endif + copy(p,q,sizeofW(StgThunk)+2,stp); + return; + + case FUN_1_1: + case FUN_2_0: + case FUN_0_2: + case CONSTR_1_1: + case CONSTR_2_0: + copy_tag(p,q,sizeofW(StgHeader)+2,stp,tag); + return; + + case CONSTR_0_2: + copy_tag(p,q,sizeofW(StgHeader)+2,stp,tag); + return; + + case THUNK: + copy(p,q,thunk_sizeW_fromITBL(info),stp); + return; + + case FUN: + case IND_PERM: + case IND_OLDGEN_PERM: + case WEAK: + case STABLE_NAME: + case CONSTR: + copy_tag(p,q,sizeW_fromITBL(info),stp,tag); + return; + + case BCO: + copy(p,q,bco_sizeW((StgBCO *)q),stp); + return; + + case CAF_BLACKHOLE: + case SE_CAF_BLACKHOLE: + case SE_BLACKHOLE: + case BLACKHOLE: + copyPart(p,q,BLACKHOLE_sizeW(),sizeofW(StgHeader),stp); + return; + + case THUNK_SELECTOR: + eval_thunk_selector(p, (StgSelector *)q, rtsTrue); + return; + + case IND: + case IND_OLDGEN: + // follow chains of indirections, don't evacuate them + q = ((StgInd*)q)->indirectee; + *p = q; + goto loop; + + case RET_BCO: + case RET_SMALL: + case RET_BIG: + case RET_DYN: + case UPDATE_FRAME: + case STOP_FRAME: + case CATCH_FRAME: + case CATCH_STM_FRAME: + case CATCH_RETRY_FRAME: + case ATOMICALLY_FRAME: + // shouldn't see these + barf("evacuate: stack frame at %p\n", q); + + case PAP: + copy(p,q,pap_sizeW((StgPAP*)q),stp); + return; + + case AP: + copy(p,q,ap_sizeW((StgAP*)q),stp); + return; + + case AP_STACK: + copy(p,q,ap_stack_sizeW((StgAP_STACK*)q),stp); + return; + + case EVACUATED: + /* Already evacuated, just return the forwarding address. + * HOWEVER: if the requested destination generation (gct->evac_step) is + * older than the actual generation (because the object was + * already evacuated to a younger generation) then we have to + * set the gct->failed_to_evac flag to indicate that we couldn't + * manage to promote the object to the desired generation. + */ + /* + * Optimisation: the check is fairly expensive, but we can often + * shortcut it if either the required generation is 0, or the + * current object (the EVACUATED) is in a high enough generation. + * We know that an EVACUATED always points to an object in the + * same or an older generation. stp is the lowest step that the + * current object would be evacuated to, so we only do the full + * check if stp is too low. + */ + { + StgClosure *e = ((StgEvacuated*)q)->evacuee; + *p = e; + if (stp < gct->evac_step) { // optimisation + if (HEAP_ALLOCED(e) && Bdescr((P_)e)->step < gct->evac_step) { + gct->failed_to_evac = rtsTrue; + TICK_GC_FAILED_PROMOTION(); + } + } + return; + } + + case ARR_WORDS: + // just copy the block + copy(p,q,arr_words_sizeW((StgArrWords *)q),stp); + return; + + case MUT_ARR_PTRS_CLEAN: + case MUT_ARR_PTRS_DIRTY: + case MUT_ARR_PTRS_FROZEN: + case MUT_ARR_PTRS_FROZEN0: + // just copy the block + copy(p,q,mut_arr_ptrs_sizeW((StgMutArrPtrs *)q),stp); + return; + + case TSO: + { + StgTSO *tso = (StgTSO *)q; + + /* Deal with redirected TSOs (a TSO that's had its stack enlarged). + */ + if (tso->what_next == ThreadRelocated) { + q = (StgClosure *)tso->link; + *p = q; + goto loop; + } + + /* To evacuate a small TSO, we need to relocate the update frame + * list it contains. + */ + { + StgTSO *new_tso; + StgPtr r, s; + + copyPart(p,(StgClosure *)tso, tso_sizeW(tso), sizeofW(StgTSO), stp); + new_tso = (StgTSO *)*p; + move_TSO(tso, new_tso); + for (r = tso->sp, s = new_tso->sp; + r < tso->stack+tso->stack_size;) { + *s++ = *r++; + } + return; + } + } + + case TREC_HEADER: + copy(p,q,sizeofW(StgTRecHeader),stp); + return; + + case TVAR_WATCH_QUEUE: + copy(p,q,sizeofW(StgTVarWatchQueue),stp); + return; + + case TVAR: + copy(p,q,sizeofW(StgTVar),stp); + return; + + case TREC_CHUNK: + copy(p,q,sizeofW(StgTRecChunk),stp); + return; + + case ATOMIC_INVARIANT: + copy(p,q,sizeofW(StgAtomicInvariant),stp); + return; + + case INVARIANT_CHECK_QUEUE: + copy(p,q,sizeofW(StgInvariantCheckQueue),stp); + return; + + default: + barf("evacuate: strange closure type %d", (int)(info->type)); + } + + barf("evacuate"); +} diff --git a/rts/sm/Evac.h b/rts/sm/Evac.h index c73008f..33ee4c5 100644 --- a/rts/sm/Evac.h +++ b/rts/sm/Evac.h @@ -27,6 +27,7 @@ #define REGPARM1 #endif -REGPARM1 void evacuate (StgClosure **p); +REGPARM1 void evacuate (StgClosure **p); +REGPARM1 void evacuate0 (StgClosure **p); extern lnat thunk_selector_depth; diff --git a/rts/sm/Scav.c b/rts/sm/Scav.c index 71c2be7..17e519d 100644 --- a/rts/sm/Scav.c +++ b/rts/sm/Scav.c @@ -31,8 +31,6 @@ static void scavenge_large_bitmap (StgPtr p, StgLargeBitmap *large_bitmap, nat size ); -static void scavenge_block (bdescr *bd, StgPtr scan); - /* Similar to scavenge_large_bitmap(), but we don't write back the * pointers we get back from evacuate(). @@ -252,429 +250,6 @@ scavenge_AP (StgAP *ap) } /* ----------------------------------------------------------------------------- - Scavenge a block from the given scan pointer up to bd->free. - - evac_step is set by the caller to be either zero (for a step in a - generation < N) or G where G is the generation of the step being - scavenged. - - We sometimes temporarily change evac_step back to zero if we're - scavenging a mutable object where eager promotion isn't such a good - idea. - -------------------------------------------------------------------------- */ - -static void -scavenge_block (bdescr *bd, StgPtr scan) -{ - StgPtr p, q; - StgInfoTable *info; - step *saved_evac_step; - - p = scan; - - debugTrace(DEBUG_gc, "scavenging block %p (gen %d, step %d) @ %p", - bd->start, bd->gen_no, bd->step->no, scan); - - gct->evac_step = bd->step; - saved_evac_step = gct->evac_step; - gct->failed_to_evac = rtsFalse; - - // we might be evacuating into the very object that we're - // scavenging, so we have to check the real bd->free pointer each - // time around the loop. - while (p < bd->free) { - - ASSERT(LOOKS_LIKE_CLOSURE_PTR(p)); - info = get_itbl((StgClosure *)p); - - ASSERT(gct->thunk_selector_depth == 0); - - q = p; - switch (info->type) { - - case MVAR_CLEAN: - case MVAR_DIRTY: - { - rtsBool saved_eager_promotion = gct->eager_promotion; - - StgMVar *mvar = ((StgMVar *)p); - gct->eager_promotion = rtsFalse; - evacuate((StgClosure **)&mvar->head); - evacuate((StgClosure **)&mvar->tail); - evacuate((StgClosure **)&mvar->value); - gct->eager_promotion = saved_eager_promotion; - - if (gct->failed_to_evac) { - mvar->header.info = &stg_MVAR_DIRTY_info; - } else { - mvar->header.info = &stg_MVAR_CLEAN_info; - } - p += sizeofW(StgMVar); - break; - } - - case FUN_2_0: - scavenge_fun_srt(info); - evacuate(&((StgClosure *)p)->payload[1]); - evacuate(&((StgClosure *)p)->payload[0]); - p += sizeofW(StgHeader) + 2; - break; - - case THUNK_2_0: - scavenge_thunk_srt(info); - evacuate(&((StgThunk *)p)->payload[1]); - evacuate(&((StgThunk *)p)->payload[0]); - p += sizeofW(StgThunk) + 2; - break; - - case CONSTR_2_0: - evacuate(&((StgClosure *)p)->payload[1]); - evacuate(&((StgClosure *)p)->payload[0]); - p += sizeofW(StgHeader) + 2; - break; - - case THUNK_1_0: - scavenge_thunk_srt(info); - evacuate(&((StgThunk *)p)->payload[0]); - p += sizeofW(StgThunk) + 1; - break; - - case FUN_1_0: - scavenge_fun_srt(info); - case CONSTR_1_0: - evacuate(&((StgClosure *)p)->payload[0]); - p += sizeofW(StgHeader) + 1; - break; - - case THUNK_0_1: - scavenge_thunk_srt(info); - p += sizeofW(StgThunk) + 1; - break; - - case FUN_0_1: - scavenge_fun_srt(info); - case CONSTR_0_1: - p += sizeofW(StgHeader) + 1; - break; - - case THUNK_0_2: - scavenge_thunk_srt(info); - p += sizeofW(StgThunk) + 2; - break; - - case FUN_0_2: - scavenge_fun_srt(info); - case CONSTR_0_2: - p += sizeofW(StgHeader) + 2; - break; - - case THUNK_1_1: - scavenge_thunk_srt(info); - evacuate(&((StgThunk *)p)->payload[0]); - p += sizeofW(StgThunk) + 2; - break; - - case FUN_1_1: - scavenge_fun_srt(info); - case CONSTR_1_1: - evacuate(&((StgClosure *)p)->payload[0]); - p += sizeofW(StgHeader) + 2; - break; - - case FUN: - scavenge_fun_srt(info); - goto gen_obj; - - case THUNK: - { - StgPtr end; - - scavenge_thunk_srt(info); - end = (P_)((StgThunk *)p)->payload + info->layout.payload.ptrs; - for (p = (P_)((StgThunk *)p)->payload; p < end; p++) { - evacuate((StgClosure **)p); - } - p += info->layout.payload.nptrs; - break; - } - - gen_obj: - case CONSTR: - case WEAK: - case STABLE_NAME: - { - StgPtr end; - - end = (P_)((StgClosure *)p)->payload + info->layout.payload.ptrs; - for (p = (P_)((StgClosure *)p)->payload; p < end; p++) { - evacuate((StgClosure **)p); - } - p += info->layout.payload.nptrs; - break; - } - - case BCO: { - StgBCO *bco = (StgBCO *)p; - evacuate((StgClosure **)&bco->instrs); - evacuate((StgClosure **)&bco->literals); - evacuate((StgClosure **)&bco->ptrs); - p += bco_sizeW(bco); - break; - } - - case IND_PERM: - if (bd->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_RECORD_CREATE()? - // - SET_INFO(((StgClosure *)p), &stg_IND_OLDGEN_PERM_info); - - // We pretend that p has just been created. - LDV_RECORD_CREATE((StgClosure *)p); - } - // fall through - case IND_OLDGEN_PERM: - evacuate(&((StgInd *)p)->indirectee); - p += sizeofW(StgInd); - break; - - case MUT_VAR_CLEAN: - case MUT_VAR_DIRTY: { - rtsBool saved_eager_promotion = gct->eager_promotion; - - gct->eager_promotion = rtsFalse; - evacuate(&((StgMutVar *)p)->var); - gct->eager_promotion = saved_eager_promotion; - - if (gct->failed_to_evac) { - ((StgClosure *)q)->header.info = &stg_MUT_VAR_DIRTY_info; - } else { - ((StgClosure *)q)->header.info = &stg_MUT_VAR_CLEAN_info; - } - p += sizeofW(StgMutVar); - break; - } - - case CAF_BLACKHOLE: - case SE_CAF_BLACKHOLE: - case SE_BLACKHOLE: - case BLACKHOLE: - p += BLACKHOLE_sizeW(); - break; - - case THUNK_SELECTOR: - { - StgSelector *s = (StgSelector *)p; - evacuate(&s->selectee); - p += THUNK_SELECTOR_sizeW(); - break; - } - - // A chunk of stack saved in a heap object - case AP_STACK: - { - StgAP_STACK *ap = (StgAP_STACK *)p; - - evacuate(&ap->fun); - scavenge_stack((StgPtr)ap->payload, (StgPtr)ap->payload + ap->size); - p = (StgPtr)ap->payload + ap->size; - break; - } - - case PAP: - p = scavenge_PAP((StgPAP *)p); - break; - - case AP: - p = scavenge_AP((StgAP *)p); - break; - - case ARR_WORDS: - // nothing to follow - p += arr_words_sizeW((StgArrWords *)p); - break; - - case MUT_ARR_PTRS_CLEAN: - case MUT_ARR_PTRS_DIRTY: - // follow everything - { - StgPtr next; - rtsBool saved_eager; - - // We don't eagerly promote objects pointed to by a mutable - // array, but if we find the array only points to objects in - // the same or an older generation, we mark it "clean" and - // avoid traversing it during minor GCs. - saved_eager = gct->eager_promotion; - gct->eager_promotion = rtsFalse; - next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p); - for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) { - evacuate((StgClosure **)p); - } - gct->eager_promotion = saved_eager; - - if (gct->failed_to_evac) { - ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_DIRTY_info; - } else { - ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_CLEAN_info; - } - - gct->failed_to_evac = rtsTrue; // always put it on the mutable list. - break; - } - - case MUT_ARR_PTRS_FROZEN: - case MUT_ARR_PTRS_FROZEN0: - // follow everything - { - StgPtr next; - - next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p); - for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) { - evacuate((StgClosure **)p); - } - - // If we're going to put this object on the mutable list, then - // set its info ptr to MUT_ARR_PTRS_FROZEN0 to indicate that. - if (gct->failed_to_evac) { - ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN0_info; - } else { - ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN_info; - } - break; - } - - case TSO: - { - StgTSO *tso = (StgTSO *)p; - rtsBool saved_eager = gct->eager_promotion; - - gct->eager_promotion = rtsFalse; - scavengeTSO(tso); - gct->eager_promotion = saved_eager; - - if (gct->failed_to_evac) { - tso->flags |= TSO_DIRTY; - } else { - tso->flags &= ~TSO_DIRTY; - } - - gct->failed_to_evac = rtsTrue; // always on the mutable list - p += tso_sizeW(tso); - break; - } - - case TVAR_WATCH_QUEUE: - { - StgTVarWatchQueue *wq = ((StgTVarWatchQueue *) p); - gct->evac_step = 0; - evacuate((StgClosure **)&wq->closure); - evacuate((StgClosure **)&wq->next_queue_entry); - evacuate((StgClosure **)&wq->prev_queue_entry); - gct->evac_step = saved_evac_step; - gct->failed_to_evac = rtsTrue; // mutable - p += sizeofW(StgTVarWatchQueue); - break; - } - - case TVAR: - { - StgTVar *tvar = ((StgTVar *) p); - gct->evac_step = 0; - evacuate((StgClosure **)&tvar->current_value); - evacuate((StgClosure **)&tvar->first_watch_queue_entry); - gct->evac_step = saved_evac_step; - gct->failed_to_evac = rtsTrue; // mutable - p += sizeofW(StgTVar); - break; - } - - case TREC_HEADER: - { - StgTRecHeader *trec = ((StgTRecHeader *) p); - gct->evac_step = 0; - evacuate((StgClosure **)&trec->enclosing_trec); - evacuate((StgClosure **)&trec->current_chunk); - evacuate((StgClosure **)&trec->invariants_to_check); - gct->evac_step = saved_evac_step; - gct->failed_to_evac = rtsTrue; // mutable - p += sizeofW(StgTRecHeader); - break; - } - - case TREC_CHUNK: - { - StgWord i; - StgTRecChunk *tc = ((StgTRecChunk *) p); - TRecEntry *e = &(tc -> entries[0]); - gct->evac_step = 0; - evacuate((StgClosure **)&tc->prev_chunk); - for (i = 0; i < tc -> next_entry_idx; i ++, e++ ) { - evacuate((StgClosure **)&e->tvar); - evacuate((StgClosure **)&e->expected_value); - evacuate((StgClosure **)&e->new_value); - } - gct->evac_step = saved_evac_step; - gct->failed_to_evac = rtsTrue; // mutable - p += sizeofW(StgTRecChunk); - break; - } - - case ATOMIC_INVARIANT: - { - StgAtomicInvariant *invariant = ((StgAtomicInvariant *) p); - gct->evac_step = 0; - evacuate(&invariant->code); - evacuate((StgClosure **)&invariant->last_execution); - gct->evac_step = saved_evac_step; - gct->failed_to_evac = rtsTrue; // mutable - p += sizeofW(StgAtomicInvariant); - break; - } - - case INVARIANT_CHECK_QUEUE: - { - StgInvariantCheckQueue *queue = ((StgInvariantCheckQueue *) p); - gct->evac_step = 0; - evacuate((StgClosure **)&queue->invariant); - evacuate((StgClosure **)&queue->my_execution); - evacuate((StgClosure **)&queue->next_queue_entry); - gct->evac_step = saved_evac_step; - gct->failed_to_evac = rtsTrue; // mutable - p += sizeofW(StgInvariantCheckQueue); - break; - } - - default: - barf("scavenge: unimplemented/strange closure type %d @ %p", - info->type, p); - } - - /* - * We need to record the current object on the mutable list if - * (a) It is actually mutable, or - * (b) It contains pointers to a younger generation. - * Case (b) arises if we didn't manage to promote everything that - * the current object points to into the current generation. - */ - if (gct->failed_to_evac) { - gct->failed_to_evac = rtsFalse; - if (bd->gen_no > 0) { - recordMutableGen_GC((StgClosure *)q, &generations[bd->gen_no]); - } - } - } - - debugTrace(DEBUG_gc, " scavenged %ld bytes", (bd->free - scan) * sizeof(W_)); -} - -/* ----------------------------------------------------------------------------- Scavenge everything on the mark stack. This is slightly different from scavenge(): @@ -1813,6 +1388,15 @@ scavenge_large (step_workspace *ws) } /* ---------------------------------------------------------------------------- + Scavenge a block + ------------------------------------------------------------------------- */ + +#define MINOR_GC +#include "Scav.c-inc" +#undef MINOR_GC +#include "Scav.c-inc" + +/* ---------------------------------------------------------------------------- Find the oldest full block to scavenge, and scavenge it. ------------------------------------------------------------------------- */ @@ -1843,7 +1427,11 @@ scavenge_find_global_work (void) // to scavenge the whole thing and then push it on // our scavd list. This saves pushing out the // scan_bd block, which might be partial. - scavenge_block(bd, bd->start); + if (N == 0) { + scavenge_block0(bd, bd->start); + } else { + scavenge_block(bd, bd->start); + } push_scan_block(bd, ws); return rtsTrue; } @@ -1893,7 +1481,11 @@ scavenge_find_local_work (void) // scavenge everything up to the free pointer. if (ws->scan != NULL && ws->scan < ws->scan_bd->free) { - scavenge_block(ws->scan_bd, ws->scan); + if (N == 0) { + scavenge_block0(ws->scan_bd, ws->scan); + } else { + scavenge_block(ws->scan_bd, ws->scan); + } ws->scan = ws->scan_bd->free; flag = rtsTrue; } diff --git a/rts/sm/Scav.c-inc b/rts/sm/Scav.c-inc new file mode 100644 index 0000000..50c4088 --- /dev/null +++ b/rts/sm/Scav.c-inc @@ -0,0 +1,472 @@ +/* ----------------------------------------------------------------------------- + * + * (c) The GHC Team 1998-2006 + * + * Generational garbage collector: scavenging functions + * + * Documentation on the architecture of the Garbage Collector can be + * found in the online commentary: + * + * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage/GC + * + * ---------------------------------------------------------------------------*/ + +// This file is #included into Scav.c, twice: firstly with MINOR_GC +// defined, the second time without. + +#ifdef MINOR_GC +#define scavenge_block(a,b) scavenge_block0(a,b) +#define evacuate(a) evacuate0(a) +#else +#undef scavenge_block +#undef evacuate +#endif + +static void scavenge_block (bdescr *bd, StgPtr scan); + +/* ----------------------------------------------------------------------------- + Scavenge a block from the given scan pointer up to bd->free. + + evac_step is set by the caller to be either zero (for a step in a + generation < N) or G where G is the generation of the step being + scavenged. + + We sometimes temporarily change evac_step back to zero if we're + scavenging a mutable object where eager promotion isn't such a good + idea. + -------------------------------------------------------------------------- */ + +static void +scavenge_block (bdescr *bd, StgPtr scan) +{ + StgPtr p, q; + StgInfoTable *info; + step *saved_evac_step; + + p = scan; + + debugTrace(DEBUG_gc, "scavenging block %p (gen %d, step %d) @ %p", + bd->start, bd->gen_no, bd->step->no, scan); + + gct->evac_step = bd->step; + saved_evac_step = gct->evac_step; + gct->failed_to_evac = rtsFalse; + + // we might be evacuating into the very object that we're + // scavenging, so we have to check the real bd->free pointer each + // time around the loop. + while (p < bd->free) { + + ASSERT(LOOKS_LIKE_CLOSURE_PTR(p)); + info = get_itbl((StgClosure *)p); + + ASSERT(gct->thunk_selector_depth == 0); + + q = p; + switch (info->type) { + + case MVAR_CLEAN: + case MVAR_DIRTY: + { + rtsBool saved_eager_promotion = gct->eager_promotion; + + StgMVar *mvar = ((StgMVar *)p); + gct->eager_promotion = rtsFalse; + evacuate((StgClosure **)&mvar->head); + evacuate((StgClosure **)&mvar->tail); + evacuate((StgClosure **)&mvar->value); + gct->eager_promotion = saved_eager_promotion; + + if (gct->failed_to_evac) { + mvar->header.info = &stg_MVAR_DIRTY_info; + } else { + mvar->header.info = &stg_MVAR_CLEAN_info; + } + p += sizeofW(StgMVar); + break; + } + + case FUN_2_0: +#ifndef MINOR_GC + scavenge_fun_srt(info); +#endif + evacuate(&((StgClosure *)p)->payload[1]); + evacuate(&((StgClosure *)p)->payload[0]); + p += sizeofW(StgHeader) + 2; + break; + + case THUNK_2_0: +#ifndef MINOR_GC + scavenge_thunk_srt(info); +#endif + evacuate(&((StgThunk *)p)->payload[1]); + evacuate(&((StgThunk *)p)->payload[0]); + p += sizeofW(StgThunk) + 2; + break; + + case CONSTR_2_0: + evacuate(&((StgClosure *)p)->payload[1]); + evacuate(&((StgClosure *)p)->payload[0]); + p += sizeofW(StgHeader) + 2; + break; + + case THUNK_1_0: +#ifndef MINOR_GC + scavenge_thunk_srt(info); +#endif + evacuate(&((StgThunk *)p)->payload[0]); + p += sizeofW(StgThunk) + 1; + break; + + case FUN_1_0: +#ifndef MINOR_GC + scavenge_fun_srt(info); +#endif + case CONSTR_1_0: + evacuate(&((StgClosure *)p)->payload[0]); + p += sizeofW(StgHeader) + 1; + break; + + case THUNK_0_1: +#ifndef MINOR_GC + scavenge_thunk_srt(info); +#endif + p += sizeofW(StgThunk) + 1; + break; + + case FUN_0_1: +#ifndef MINOR_GC + scavenge_fun_srt(info); +#endif + case CONSTR_0_1: + p += sizeofW(StgHeader) + 1; + break; + + case THUNK_0_2: +#ifndef MINOR_GC + scavenge_thunk_srt(info); +#endif + p += sizeofW(StgThunk) + 2; + break; + + case FUN_0_2: +#ifndef MINOR_GC + scavenge_fun_srt(info); +#endif + case CONSTR_0_2: + p += sizeofW(StgHeader) + 2; + break; + + case THUNK_1_1: +#ifndef MINOR_GC + scavenge_thunk_srt(info); +#endif + evacuate(&((StgThunk *)p)->payload[0]); + p += sizeofW(StgThunk) + 2; + break; + + case FUN_1_1: +#ifndef MINOR_GC + scavenge_fun_srt(info); +#endif + case CONSTR_1_1: + evacuate(&((StgClosure *)p)->payload[0]); + p += sizeofW(StgHeader) + 2; + break; + + case FUN: +#ifndef MINOR_GC + scavenge_fun_srt(info); +#endif + goto gen_obj; + + case THUNK: + { + StgPtr end; + +#ifndef MINOR_GC + scavenge_thunk_srt(info); +#endif + end = (P_)((StgThunk *)p)->payload + info->layout.payload.ptrs; + for (p = (P_)((StgThunk *)p)->payload; p < end; p++) { + evacuate((StgClosure **)p); + } + p += info->layout.payload.nptrs; + break; + } + + gen_obj: + case CONSTR: + case WEAK: + case STABLE_NAME: + { + StgPtr end; + + end = (P_)((StgClosure *)p)->payload + info->layout.payload.ptrs; + for (p = (P_)((StgClosure *)p)->payload; p < end; p++) { + evacuate((StgClosure **)p); + } + p += info->layout.payload.nptrs; + break; + } + + case BCO: { + StgBCO *bco = (StgBCO *)p; + evacuate((StgClosure **)&bco->instrs); + evacuate((StgClosure **)&bco->literals); + evacuate((StgClosure **)&bco->ptrs); + p += bco_sizeW(bco); + break; + } + + case IND_PERM: + if (bd->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_RECORD_CREATE()? + // + SET_INFO(((StgClosure *)p), &stg_IND_OLDGEN_PERM_info); + + // We pretend that p has just been created. + LDV_RECORD_CREATE((StgClosure *)p); + } + // fall through + case IND_OLDGEN_PERM: + evacuate(&((StgInd *)p)->indirectee); + p += sizeofW(StgInd); + break; + + case MUT_VAR_CLEAN: + case MUT_VAR_DIRTY: { + rtsBool saved_eager_promotion = gct->eager_promotion; + + gct->eager_promotion = rtsFalse; + evacuate(&((StgMutVar *)p)->var); + gct->eager_promotion = saved_eager_promotion; + + if (gct->failed_to_evac) { + ((StgClosure *)q)->header.info = &stg_MUT_VAR_DIRTY_info; + } else { + ((StgClosure *)q)->header.info = &stg_MUT_VAR_CLEAN_info; + } + p += sizeofW(StgMutVar); + break; + } + + case CAF_BLACKHOLE: + case SE_CAF_BLACKHOLE: + case SE_BLACKHOLE: + case BLACKHOLE: + p += BLACKHOLE_sizeW(); + break; + + case THUNK_SELECTOR: + { + StgSelector *s = (StgSelector *)p; + evacuate(&s->selectee); + p += THUNK_SELECTOR_sizeW(); + break; + } + + // A chunk of stack saved in a heap object + case AP_STACK: + { + StgAP_STACK *ap = (StgAP_STACK *)p; + + evacuate(&ap->fun); + scavenge_stack((StgPtr)ap->payload, (StgPtr)ap->payload + ap->size); + p = (StgPtr)ap->payload + ap->size; + break; + } + + case PAP: + p = scavenge_PAP((StgPAP *)p); + break; + + case AP: + p = scavenge_AP((StgAP *)p); + break; + + case ARR_WORDS: + // nothing to follow + p += arr_words_sizeW((StgArrWords *)p); + break; + + case MUT_ARR_PTRS_CLEAN: + case MUT_ARR_PTRS_DIRTY: + // follow everything + { + StgPtr next; + rtsBool saved_eager; + + // We don't eagerly promote objects pointed to by a mutable + // array, but if we find the array only points to objects in + // the same or an older generation, we mark it "clean" and + // avoid traversing it during minor GCs. + saved_eager = gct->eager_promotion; + gct->eager_promotion = rtsFalse; + next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p); + for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) { + evacuate((StgClosure **)p); + } + gct->eager_promotion = saved_eager; + + if (gct->failed_to_evac) { + ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_DIRTY_info; + } else { + ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_CLEAN_info; + } + + gct->failed_to_evac = rtsTrue; // always put it on the mutable list. + break; + } + + case MUT_ARR_PTRS_FROZEN: + case MUT_ARR_PTRS_FROZEN0: + // follow everything + { + StgPtr next; + + next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p); + for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) { + evacuate((StgClosure **)p); + } + + // If we're going to put this object on the mutable list, then + // set its info ptr to MUT_ARR_PTRS_FROZEN0 to indicate that. + if (gct->failed_to_evac) { + ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN0_info; + } else { + ((StgClosure *)q)->header.info = &stg_MUT_ARR_PTRS_FROZEN_info; + } + break; + } + + case TSO: + { + StgTSO *tso = (StgTSO *)p; + rtsBool saved_eager = gct->eager_promotion; + + gct->eager_promotion = rtsFalse; + scavengeTSO(tso); + gct->eager_promotion = saved_eager; + + if (gct->failed_to_evac) { + tso->flags |= TSO_DIRTY; + } else { + tso->flags &= ~TSO_DIRTY; + } + + gct->failed_to_evac = rtsTrue; // always on the mutable list + p += tso_sizeW(tso); + break; + } + + case TVAR_WATCH_QUEUE: + { + StgTVarWatchQueue *wq = ((StgTVarWatchQueue *) p); + gct->evac_step = 0; + evacuate((StgClosure **)&wq->closure); + evacuate((StgClosure **)&wq->next_queue_entry); + evacuate((StgClosure **)&wq->prev_queue_entry); + gct->evac_step = saved_evac_step; + gct->failed_to_evac = rtsTrue; // mutable + p += sizeofW(StgTVarWatchQueue); + break; + } + + case TVAR: + { + StgTVar *tvar = ((StgTVar *) p); + gct->evac_step = 0; + evacuate((StgClosure **)&tvar->current_value); + evacuate((StgClosure **)&tvar->first_watch_queue_entry); + gct->evac_step = saved_evac_step; + gct->failed_to_evac = rtsTrue; // mutable + p += sizeofW(StgTVar); + break; + } + + case TREC_HEADER: + { + StgTRecHeader *trec = ((StgTRecHeader *) p); + gct->evac_step = 0; + evacuate((StgClosure **)&trec->enclosing_trec); + evacuate((StgClosure **)&trec->current_chunk); + evacuate((StgClosure **)&trec->invariants_to_check); + gct->evac_step = saved_evac_step; + gct->failed_to_evac = rtsTrue; // mutable + p += sizeofW(StgTRecHeader); + break; + } + + case TREC_CHUNK: + { + StgWord i; + StgTRecChunk *tc = ((StgTRecChunk *) p); + TRecEntry *e = &(tc -> entries[0]); + gct->evac_step = 0; + evacuate((StgClosure **)&tc->prev_chunk); + for (i = 0; i < tc -> next_entry_idx; i ++, e++ ) { + evacuate((StgClosure **)&e->tvar); + evacuate((StgClosure **)&e->expected_value); + evacuate((StgClosure **)&e->new_value); + } + gct->evac_step = saved_evac_step; + gct->failed_to_evac = rtsTrue; // mutable + p += sizeofW(StgTRecChunk); + break; + } + + case ATOMIC_INVARIANT: + { + StgAtomicInvariant *invariant = ((StgAtomicInvariant *) p); + gct->evac_step = 0; + evacuate(&invariant->code); + evacuate((StgClosure **)&invariant->last_execution); + gct->evac_step = saved_evac_step; + gct->failed_to_evac = rtsTrue; // mutable + p += sizeofW(StgAtomicInvariant); + break; + } + + case INVARIANT_CHECK_QUEUE: + { + StgInvariantCheckQueue *queue = ((StgInvariantCheckQueue *) p); + gct->evac_step = 0; + evacuate((StgClosure **)&queue->invariant); + evacuate((StgClosure **)&queue->my_execution); + evacuate((StgClosure **)&queue->next_queue_entry); + gct->evac_step = saved_evac_step; + gct->failed_to_evac = rtsTrue; // mutable + p += sizeofW(StgInvariantCheckQueue); + break; + } + + default: + barf("scavenge: unimplemented/strange closure type %d @ %p", + info->type, p); + } + + /* + * We need to record the current object on the mutable list if + * (a) It is actually mutable, or + * (b) It contains pointers to a younger generation. + * Case (b) arises if we didn't manage to promote everything that + * the current object points to into the current generation. + */ + if (gct->failed_to_evac) { + gct->failed_to_evac = rtsFalse; + if (bd->gen_no > 0) { + recordMutableGen_GC((StgClosure *)q, &generations[bd->gen_no]); + } + } + } + + debugTrace(DEBUG_gc, " scavenged %ld bytes", (bd->free - scan) * sizeof(W_)); +} -- 1.7.10.4