X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;ds=sidebyside;f=ghc%2Frts%2FGC.c;h=7578d1c1f41974e20a24cd7f693f7d5bc54749e1;hb=39a991713f0705bf2599930e378178a1d6e2b5f9;hp=fa52ddaa90429ecb85dc611b9c46049dfb5f4378;hpb=723545930025a24708a8a0923435c95cc7f058c9;p=ghc-hetmet.git diff --git a/ghc/rts/GC.c b/ghc/rts/GC.c index fa52dda..7578d1c 100644 --- a/ghc/rts/GC.c +++ b/ghc/rts/GC.c @@ -1,10 +1,31 @@ /* ----------------------------------------------------------------------------- - * $Id: GC.c,v 1.20 1999/01/26 16:16:22 simonm Exp $ + * $Id: GC.c,v 1.86 2000/11/01 11:41:47 simonmar Exp $ * - * Two-space garbage collector + * (c) The GHC Team 1998-1999 + * + * Generational garbage collector * * ---------------------------------------------------------------------------*/ +//@menu +//* Includes:: +//* STATIC OBJECT LIST:: +//* Static function declarations:: +//* Garbage Collect:: +//* Weak Pointers:: +//* Evacuation:: +//* Scavenging:: +//* Reverting CAFs:: +//* Sanity code for CAF garbage collection:: +//* Lazy black holing:: +//* Stack squeezing:: +//* Pausing a thread:: +//* Index:: +//@end menu + +//@node Includes, STATIC OBJECT LIST +//@subsection Includes + #include "Rts.h" #include "RtsFlags.h" #include "RtsUtils.h" @@ -17,12 +38,30 @@ #include "GC.h" #include "BlockAlloc.h" #include "Main.h" -#include "DebugProf.h" +#include "ProfHeap.h" #include "SchedAPI.h" #include "Weak.h" #include "StablePriv.h" +#include "Prelude.h" +#if defined(GRAN) || defined(PAR) +# include "GranSimRts.h" +# include "ParallelRts.h" +# include "FetchMe.h" +# if defined(DEBUG) +# include "Printer.h" +# include "ParallelDebug.h" +# endif +#endif +#if defined(GHCI) +# include "HsFFI.h" +# include "Linker.h" +#endif +#if defined(RTS_GTK_FRONTPANEL) +#include "FrontPanel.h" +#endif -StgCAF* enteredCAFs; +//@node STATIC OBJECT LIST, Static function declarations, Includes +//@subsection STATIC OBJECT LIST /* STATIC OBJECT LIST. * @@ -75,11 +114,16 @@ static rtsBool major_gc; */ static nat evac_gen; -/* WEAK POINTERS +/* Weak pointers */ static StgWeak *old_weak_ptr_list; /* also pending finaliser list */ static rtsBool weak_done; /* all done for this pass */ +/* List of all threads during GC + */ +static StgTSO *old_all_threads; +static StgTSO *resurrected_threads; + /* Flag indicating failure to evacuate an object to the desired * generation. */ @@ -89,26 +133,39 @@ static rtsBool failed_to_evac; */ bdescr *old_to_space; +/* 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 */ + +//@node Static function declarations, Garbage Collect, STATIC OBJECT LIST +//@subsection Static function declarations + /* ----------------------------------------------------------------------------- Static function declarations -------------------------------------------------------------------------- */ -static StgClosure *evacuate(StgClosure *q); -static void zeroStaticObjectList(StgClosure* first_static); -static rtsBool traverse_weak_ptr_list(void); -static void zeroMutableList(StgMutClosure *first); -static void revertDeadCAFs(void); +static StgClosure * evacuate ( StgClosure *q ); +static void zero_static_object_list ( StgClosure* first_static ); +static void zero_mutable_list ( StgMutClosure *first ); -static void scavenge_stack(StgPtr p, StgPtr stack_end); -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 rtsBool traverse_weak_ptr_list ( void ); +static void cleanup_weak_ptr_list ( StgWeak **list ); + +static void scavenge_stack ( StgPtr p, StgPtr stack_end ); +static void scavenge_large ( step *step ); +static void scavenge ( step *step ); +static void scavenge_static ( void ); +static void scavenge_mutable_list ( generation *g ); +static void scavenge_mut_once_list ( generation *g ); #ifdef DEBUG -static void gcCAFs(void); +static void gcCAFs ( void ); #endif +//@node Garbage Collect, Weak Pointers, Static function declarations +//@subsection Garbage Collect + /* ----------------------------------------------------------------------------- GarbageCollect @@ -131,18 +188,24 @@ static void gcCAFs(void); - free from-space in each step, and set from-space = to-space. -------------------------------------------------------------------------- */ +//@cindex GarbageCollect -void GarbageCollect(void (*get_roots)(void)) +void GarbageCollect ( void (*get_roots)(void), rtsBool force_major_gc ) { bdescr *bd; step *step; - lnat live, allocated, collected = 0; + lnat live, allocated, collected = 0, copied = 0; nat g, s; #ifdef PROFILING CostCentreStack *prev_CCS; #endif +#if defined(DEBUG) && defined(GRAN) + IF_DEBUG(gc, belch("@@ Starting garbage collection at %ld (%lx)\n", + Now, Now)); +#endif + /* tell the stats department that we've started a GC */ stat_startGC(); @@ -152,34 +215,36 @@ void GarbageCollect(void (*get_roots)(void)) CCCS = CCS_GC; #endif - /* We might have been called from Haskell land by _ccall_GC, in - * which case we need to call threadPaused() because the scheduler - * won't have done it. - */ - if (CurrentTSO) { threadPaused(CurrentTSO); } - - /* Approximate how much we allocated: number of blocks in the - * nursery + blocks allocated via allocate() - unused nusery blocks. - * This leaves a little slop at the end of each block, and doesn't - * take into account large objects (ToDo). + /* Approximate how much we allocated. + * Todo: only when generating stats? */ - allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes(); - for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) { - allocated -= BLOCK_SIZE_W; - } + allocated = calcAllocated(); /* 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; + if (force_major_gc) { + N = RtsFlags.GcFlags.generations - 1; + major_gc = rtsTrue; + } else { + N = 0; + for (g = 0; g < RtsFlags.GcFlags.generations; g++) { + if (generations[g].steps[0].n_blocks >= generations[g].max_blocks) { + N = g; + } } + major_gc = (N == RtsFlags.GcFlags.generations-1); + } + +#ifdef RTS_GTK_FRONTPANEL + if (RtsFlags.GcFlags.frontpanel) { + updateFrontPanelBeforeGC(N); } - major_gc = (N == RtsFlags.GcFlags.generations-1); +#endif /* check stack sanity *before* GC (ToDo: check all threads) */ - /*IF_DEBUG(sanity, checkTSO(MainTSO,0)); */ +#if defined(GRAN) + // ToDo!: check sanity IF_DEBUG(sanity, checkTSOsSanity()); +#endif IF_DEBUG(sanity, checkFreeListSanity()); /* Initialise the static object lists @@ -188,10 +253,10 @@ void GarbageCollect(void (*get_roots)(void)) scavenged_static_objects = END_OF_STATIC_LIST; /* zero the mutable list for the oldest generation (see comment by - * zeroMutableList below). + * zero_mutable_list below). */ if (major_gc) { - zeroMutableList(generations[RtsFlags.GcFlags.generations-1].mut_list); + zero_mutable_list(generations[RtsFlags.GcFlags.generations-1].mut_once_list); } /* Save the old to-space if we're doing a two-space collection @@ -201,10 +266,16 @@ void GarbageCollect(void (*get_roots)(void)) g0s0->to_space = NULL; } + /* Keep a count of how many new blocks we allocated during this GC + * (used for resizing the allocation area, later). + */ + new_blocks = 0; + /* 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++) { @@ -229,11 +300,12 @@ void GarbageCollect(void (*get_roots)(void)) step->hpLim = step->hp + BLOCK_SIZE_W; step->hp_bd = bd; step->to_space = bd; - step->to_blocks = 1; /* ???? */ + step->to_blocks = 1; step->scan = bd->start; step->scan_bd = bd; step->new_large_objects = NULL; step->scavenged_large_objects = NULL; + new_blocks++; /* mark the large objects as not evacuated yet */ for (bd = step->large_objects; bd; bd = bd->link) { bd->evacuated = 0; @@ -258,6 +330,7 @@ void GarbageCollect(void (*get_roots)(void)) step->hp_bd = bd; step->blocks = bd; step->n_blocks = 1; + new_blocks++; } /* Set the scan pointer for older generations: remember we * still have to scavenge objects that have been promoted. */ @@ -283,19 +356,31 @@ void GarbageCollect(void (*get_roots)(void)) * it has already been evaced to gen 2. */ { - StgMutClosure *tmp, **pp; - for (g = N+1; g < RtsFlags.GcFlags.generations; g++) { + int st; + for (g = RtsFlags.GcFlags.generations-1; g > N; g--) { generations[g].saved_mut_list = generations[g].mut_list; generations[g].mut_list = END_MUT_LIST; } + /* Do the mut-once lists first */ 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; + IF_PAR_DEBUG(verbose, + printMutOnceList(&generations[g])); + scavenge_mut_once_list(&generations[g]); + evac_gen = g; + for (st = generations[g].n_steps-1; st >= 0; st--) { + scavenge(&generations[g].steps[st]); + } + } + + for (g = RtsFlags.GcFlags.generations-1; g > N; g--) { + IF_PAR_DEBUG(verbose, + printMutableList(&generations[g])); + scavenge_mutable_list(&generations[g]); + evac_gen = g; + for (st = generations[g].n_steps-1; st >= 0; st--) { + scavenge(&generations[g].steps[st]); } - *pp = tmp; } } @@ -304,20 +389,33 @@ void GarbageCollect(void (*get_roots)(void)) evac_gen = 0; get_roots(); +#if defined(PAR) /* And don't forget to mark the TSO if we got here direct from * Haskell! */ + /* Not needed in a seq version? if (CurrentTSO) { CurrentTSO = (StgTSO *)MarkRoot((StgClosure *)CurrentTSO); } + */ + + /* Mark the entries in the GALA table of the parallel system */ + markLocalGAs(major_gc); +#endif /* Mark the weak pointer list, and prepare to detect dead weak * pointers. */ - markWeakList(); old_weak_ptr_list = weak_ptr_list; weak_ptr_list = NULL; weak_done = rtsFalse; + /* The all_threads list is like the weak_ptr_list. + * See traverse_weak_ptr_list() for the details. + */ + old_all_threads = all_threads; + all_threads = END_TSO_QUEUE; + resurrected_threads = END_TSO_QUEUE; + /* Mark the stable pointer table. */ markStablePtrTable(major_gc); @@ -329,22 +427,7 @@ void GarbageCollect(void (*get_roots)(void)) * the CAF document. */ extern void markHugsObjects(void); -#if 0 - /* ToDo: This (undefined) function should contain the scavenge - * loop immediately below this block of code - but I'm not sure - * enough of the details to do this myself. - */ - scavengeEverything(); - /* revert dead CAFs and update enteredCAFs list */ - revertDeadCAFs(); -#endif markHugsObjects(); -#if 0 - /* This will keep the CAFs and the attached BCOs alive - * but the values will have been reverted - */ - scavengeEverything(); -#endif } #endif @@ -359,6 +442,8 @@ void GarbageCollect(void (*get_roots)(void)) /* scavenge static objects */ if (major_gc && static_objects != END_OF_STATIC_LIST) { + IF_DEBUG(sanity, + checkStaticObjects()); scavenge_static(); } @@ -373,18 +458,24 @@ void GarbageCollect(void (*get_roots)(void)) /* scavenge each step in generations 0..maxgen */ { - int gen; + int gen, st; + loop2: for (gen = RtsFlags.GcFlags.generations-1; gen >= 0; gen--) { - for (s = 0; s < generations[gen].n_steps; s++) { - step = &generations[gen].steps[s]; + for (st = generations[gen].n_steps-1; st >= 0 ; st--) { + if (gen == 0 && st == 0 && RtsFlags.GcFlags.generations > 1) { + continue; + } + step = &generations[gen].steps[st]; evac_gen = gen; if (step->hp_bd != step->scan_bd || step->scan < step->hp) { scavenge(step); flag = rtsTrue; + goto loop2; } if (step->new_large_objects != NULL) { scavenge_large(step); flag = rtsTrue; + goto loop2; } } } @@ -397,10 +488,22 @@ void GarbageCollect(void (*get_roots)(void)) } } - /* Now see which stable names are still alive + /* Final traversal of the weak pointer list (see comment by + * cleanUpWeakPtrList below). + */ + cleanup_weak_ptr_list(&weak_ptr_list); + + /* Now see which stable names are still alive. */ gcStablePtrTable(major_gc); +#if defined(PAR) + /* Reconstruct the Global Address tables used in GUM */ + rebuildGAtables(major_gc); + IF_DEBUG(sanity, checkGlobalTSOList(rtsTrue/*check TSOs, too*/)); + IF_DEBUG(sanity, checkLAGAtable(rtsTrue/*check closures, too*/)); +#endif + /* 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. @@ -424,67 +527,11 @@ void GarbageCollect(void (*get_roots)(void)) } } } - } else { - /* 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. - */ - nat blocks = g0s0->to_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; - } - } - - if (nursery_blocks < blocks) { - IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n", - blocks)); - g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks); - } else { - bdescr *next_bd; - - IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n", - blocks)); - for (bd = g0s0->blocks; nursery_blocks > blocks; nursery_blocks--) { - next_bd = bd->link; - freeGroup(bd); - bd = next_bd; - } - g0s0->blocks = bd; - } - - g0s0->n_blocks = nursery_blocks = blocks; } /* run through all the generations/steps and tidy up */ + copied = new_blocks * BLOCK_SIZE_W; for (g = 0; g < RtsFlags.GcFlags.generations; g++) { if (g <= N) { @@ -499,6 +546,11 @@ void GarbageCollect(void (*get_roots)(void)) /* Tidy the end of the to-space chains */ step->hp_bd->free = step->hp; step->hp_bd->link = NULL; + /* stats information: how much we copied */ + if (g <= N) { + copied -= step->hp_bd->start + BLOCK_SIZE_W - + step->hp_bd->free; + } } /* for generations we collected... */ @@ -540,16 +592,16 @@ void GarbageCollect(void (*get_roots)(void)) * between the maximum size of the oldest and youngest * generations. * - * max_blocks = alloc_area_size + - * (oldgen_max_blocks - alloc_area_size) * G - * ----------------------------------------- - * oldest_gen + * max_blocks = oldgen_max_blocks * G + * ---------------------- + * oldest_gen */ if (g != 0) { - generations[g].max_blocks = - RtsFlags.GcFlags.minAllocAreaSize + - (((oldest_gen->max_blocks - RtsFlags.GcFlags.minAllocAreaSize) * g) - / (RtsFlags.GcFlags.generations-1)); +#if 0 + generations[g].max_blocks = (oldest_gen->max_blocks * g) + / (RtsFlags.GcFlags.generations-1); +#endif + generations[g].max_blocks = oldest_gen->max_blocks; } /* for older generations... */ @@ -571,96 +623,139 @@ void GarbageCollect(void (*get_roots)(void)) } } + /* Guess the amount of live data for stats. */ + live = calcLive(); + + /* Free the small objects allocated via allocate(), since this will + * all have been copied into G0S1 now. + */ + if (small_alloc_list != NULL) { + freeChain(small_alloc_list); + } + small_alloc_list = NULL; + alloc_blocks = 0; + alloc_Hp = NULL; + alloc_HpLim = NULL; + alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize; + /* 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 */ } - live = g0s0->to_blocks * BLOCK_SIZE_W + - ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_); - /* Generational collector: - * estimate the amount of live data. - */ + /* 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->to_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 { - 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_); + /* 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 obtained by finding the + * percentage of g0s0 that was live at the last minor GC. + */ + if (N == 0) { + g0s0_pcnt_kept = (new_blocks * 100) / g0s0->n_blocks; } + + /* Estimate a size for the allocation area based on the + * information available. We might end up going slightly under + * or over the suggested heap size, but we should be pretty + * close on average. + * + * Formula: suggested - needed + * ---------------------------- + * 1 + g0s0_pcnt_kept/100 + * + * where 'needed' is the amount of memory needed at the next + * collection for collecting all steps except g0s0. + */ + blocks = + (((int)RtsFlags.GcFlags.heapSizeSuggestion - (int)needed) * 100) / + (100 + (int)g0s0_pcnt_kept); + + if (blocks < (int)RtsFlags.GcFlags.minAllocAreaSize) { + blocks = RtsFlags.GcFlags.minAllocAreaSize; + } + + resizeNursery((nat)blocks); } } - /* revert dead CAFs and update enteredCAFs list */ - revertDeadCAFs(); - - /* mark the garbage collected CAFs as dead */ + /* mark the garbage collected CAFs as dead */ #ifdef DEBUG if (major_gc) { gcCAFs(); } #endif /* zero the scavenged static object list */ if (major_gc) { - zeroStaticObjectList(scavenged_static_objects); + zero_static_object_list(scavenged_static_objects); } /* Reset the nursery */ - for (bd = g0s0->blocks; bd; bd = bd->link) { - bd->free = bd->start; - ASSERT(bd->gen == g0); - ASSERT(bd->step == g0s0); - } - current_nursery = g0s0->blocks; + resetNurseries(); - /* Free the small objects allocated via allocate(), since this will - * all have been copied into G0S1 now. - */ - if (small_alloc_list != NULL) { - freeChain(small_alloc_list); - } - small_alloc_list = NULL; - alloc_blocks = 0; - alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize; - - /* start any pending finalisers */ - scheduleFinalisers(old_weak_ptr_list); + /* start any pending finalizers */ + scheduleFinalizers(old_weak_ptr_list); - /* check sanity after GC */ -#ifdef DEBUG - if (RtsFlags.GcFlags.generations == 1) { - IF_DEBUG(sanity, checkHeap(g0s0->to_space, NULL)); - IF_DEBUG(sanity, checkChain(g0s0->large_objects)); - } else { + /* send exceptions to any threads which were about to die */ + resurrectThreads(resurrected_threads); - 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)); - } - } - 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].blocks, - generations[g].steps[s].blocks->start)); - IF_DEBUG(sanity, checkChain(generations[g].steps[s].large_objects)); - } - } - IF_DEBUG(sanity, checkFreeListSanity()); - } -#endif + /* check sanity after GC */ + IF_DEBUG(sanity, checkSanity(N)); + /* extra GC trace info */ IF_DEBUG(gc, stat_describe_gens()); #ifdef DEBUG @@ -670,16 +765,26 @@ void GarbageCollect(void (*get_roots)(void)) /* restore enclosing cost centre */ #ifdef PROFILING + heapCensus(); CCCS = prev_CCS; #endif /* check for memory leaks if sanity checking is on */ IF_DEBUG(sanity, memInventory()); +#ifdef RTS_GTK_VISUALS + if (RtsFlags.GcFlags.visuals) { + updateFrontPanelAfterGC( N, live ); + } +#endif + /* ok, GC over: tell the stats department what happened. */ - stat_endGC(allocated, collected, live, N); + stat_endGC(allocated, collected, live, copied, N); } +//@node Weak Pointers, Evacuation, Garbage Collect +//@subsection Weak Pointers + /* ----------------------------------------------------------------------------- Weak Pointers @@ -694,11 +799,12 @@ void GarbageCollect(void (*get_roots)(void)) new live weak pointers, then all the currently unreachable ones are dead. - For generational GC: we just don't try to finalise weak pointers in + For generational GC: we just don't try to finalize weak pointers in 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. -------------------------------------------------------------------------- */ +//@cindex traverse_weak_ptr_list static rtsBool traverse_weak_ptr_list(void) @@ -709,7 +815,7 @@ traverse_weak_ptr_list(void) if (weak_done) { return rtsFalse; } - /* doesn't matter where we evacuate values/finalisers to, since + /* doesn't matter where we evacuate values/finalizers to, since * these pointers are treated as roots (iff the keys are alive). */ evac_gen = 0; @@ -717,11 +823,32 @@ traverse_weak_ptr_list(void) last_w = &old_weak_ptr_list; for (w = old_weak_ptr_list; w; w = next_w) { + /* First, this weak pointer might have been evacuated. If so, + * remove the forwarding pointer from the weak_ptr_list. + */ + if (get_itbl(w)->type == EVACUATED) { + w = (StgWeak *)((StgEvacuated *)w)->evacuee; + *last_w = 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 == &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. + */ if ((new = isAlive(w->key))) { w->key = new; - /* evacuate the value and finaliser */ + /* evacuate the value and finalizer */ w->value = evacuate(w->value); - w->finaliser = evacuate(w->finaliser); + 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 */ @@ -738,16 +865,71 @@ traverse_weak_ptr_list(void) continue; } } - + + /* 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) { + + /* Threads which have finished or died get dropped from + * the list. + */ + switch (t->what_next) { + case ThreadRelocated: + next = t->link; + *prev = next; + continue; + case ThreadKilled: + case ThreadComplete: + next = t->global_link; + *prev = next; + continue; + default: + } + + /* Threads which have already been determined to be alive are + * moved onto the all_threads list. + */ + (StgClosure *)tmp = isAlive((StgClosure *)t); + if (tmp != NULL) { + next = tmp->global_link; + tmp->global_link = all_threads; + all_threads = tmp; + *prev = next; + } else { + prev = &(t->global_link); + next = t->global_link; + } + } + } + /* 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 finalisers later on. + * of pending finalizers later on. */ if (flag == rtsFalse) { + cleanup_weak_ptr_list(&old_weak_ptr_list); for (w = old_weak_ptr_list; w; w = w->link) { - w->value = evacuate(w->value); - w->finaliser = evacuate(w->finaliser); + w->finalizer = evacuate(w->finalizer); } + + /* 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_done = rtsTrue; } @@ -755,15 +937,53 @@ traverse_weak_ptr_list(void) } /* ----------------------------------------------------------------------------- + After GC, the live weak pointer list may have forwarding pointers + on it, because a weak pointer object was evacuated after being + moved to the live weak pointer list. We remove those forwarding + pointers here. + + Also, we don't consider weak pointer objects to be reachable, but + we must nevertheless consider them to be "live" and retain them. + Therefore any weak pointer objects which haven't as yet been + evacuated need to be evacuated now. + -------------------------------------------------------------------------- */ + +//@cindex cleanup_weak_ptr_list + +static void +cleanup_weak_ptr_list ( StgWeak **list ) +{ + StgWeak *w, **last_w; + + last_w = list; + for (w = *list; w; w = w->link) { + + if (get_itbl(w)->type == EVACUATED) { + w = (StgWeak *)((StgEvacuated *)w)->evacuee; + *last_w = w; + } + + if (Bdescr((P_)w)->evacuated == 0) { + (StgClosure *)w = evacuate((StgClosure *)w); + *last_w = w; + } + last_w = &(w->link); + } +} + +/* ----------------------------------------------------------------------------- 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. -------------------------------------------------------------------------- */ +//@cindex isAlive + StgClosure * isAlive(StgClosure *p) { - StgInfoTable *info; + const StgInfoTable *info; + nat size; while (1) { @@ -794,6 +1014,33 @@ isAlive(StgClosure *p) /* alive! */ return ((StgEvacuated *)p)->evacuee; + case BCO: + size = bco_sizeW((StgBCO*)p); + goto large; + + case ARR_WORDS: + size = arr_words_sizeW((StgArrWords *)p); + goto large; + + case MUT_ARR_PTRS: + case MUT_ARR_PTRS_FROZEN: + size = mut_arr_ptrs_sizeW((StgMutArrPtrs *)p); + goto large; + + case TSO: + if (((StgTSO *)p)->what_next == ThreadRelocated) { + p = (StgClosure *)((StgTSO *)p)->link; + continue; + } + + size = tso_sizeW((StgTSO *)p); + large: + if (size >= LARGE_OBJECT_THRESHOLD/sizeof(W_) + && Bdescr((P_)p)->evacuated) + return p; + else + return NULL; + default: /* dead. */ return NULL; @@ -801,12 +1048,21 @@ isAlive(StgClosure *p) } } +//@cindex MarkRoot StgClosure * MarkRoot(StgClosure *root) { +# if 0 && defined(PAR) && defined(DEBUG) + StgClosure *foo = evacuate(root); + // ASSERT(closure_STATIC(foo) || maybeLarge(foo) || Bdescr(foo)->evacuated); + ASSERT(isAlive(foo)); // must be in to-space + return foo; +# else return evacuate(root); +# endif } +//@cindex addBlock static void addBlock(step *step) { bdescr *bd = allocBlock(); @@ -825,20 +1081,37 @@ static void addBlock(step *step) step->hpLim = step->hp + BLOCK_SIZE_W; step->hp_bd = bd; step->to_blocks++; + new_blocks++; +} + +//@cindex upd_evacuee + +static __inline__ void +upd_evacuee(StgClosure *p, StgClosure *dest) +{ + p->header.info = &EVACUATED_info; + ((StgEvacuated *)p)->evacuee = dest; } +//@cindex copy + static __inline__ StgClosure * copy(StgClosure *src, nat size, 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()). */ if (step->gen->no < evac_gen) { +#ifdef NO_EAGER_PROMOTION + failed_to_evac = rtsTrue; +#else step = &generations[evac_gen].steps[0]; +#endif } /* chain a new block onto the to-space for the destination step if @@ -854,6 +1127,7 @@ copy(StgClosure *src, nat size, step *step) dest = step->hp; step->hp = to; + upd_evacuee(src,(StgClosure *)dest); return (StgClosure *)dest; } @@ -862,13 +1136,20 @@ copy(StgClosure *src, nat size, step *step) * used to optimise evacuation of BLACKHOLEs. */ +//@cindex copyPart + 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) { +#ifdef NO_EAGER_PROMOTION + failed_to_evac = rtsTrue; +#else step = &generations[evac_gen].steps[0]; +#endif } if (step->hp + size_to_reserve >= step->hpLim) { @@ -881,36 +1162,12 @@ copyPart(StgClosure *src, nat size_to_reserve, nat size_to_copy, step *step) dest = step->hp; step->hp += size_to_reserve; + upd_evacuee(src,(StgClosure *)dest); return (StgClosure *)dest; } -static __inline__ void -upd_evacuee(StgClosure *p, StgClosure *dest) -{ - StgEvacuated *q = (StgEvacuated *)p; - - SET_INFO(q,&EVACUATED_info); - q->evacuee = 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; - } -} +//@node Evacuation, Scavenging, Weak Pointers +//@subsection Evacuation /* ----------------------------------------------------------------------------- Evacuate a large object @@ -923,6 +1180,8 @@ evacuate_mutable(StgMutClosure *c) evacuated, or 0 otherwise. -------------------------------------------------------------------------- */ +//@cindex evacuate_large + static inline void evacuate_large(StgPtr p, rtsBool mutable) { @@ -959,7 +1218,11 @@ evacuate_large(StgPtr p, rtsBool mutable) */ step = bd->step->to; if (step->gen->no < evac_gen) { +#ifdef NO_EAGER_PROMOTION + failed_to_evac = rtsTrue; +#else step = &generations[evac_gen].steps[0]; +#endif } bd->step = step; @@ -969,7 +1232,7 @@ evacuate_large(StgPtr p, rtsBool mutable) bd->evacuated = 1; if (mutable) { - evacuate_mutable((StgMutClosure *)p); + recordMutable((StgMutClosure *)p); } } @@ -981,6 +1244,8 @@ evacuate_large(StgPtr p, rtsBool mutable) the promotion until the next GC. -------------------------------------------------------------------------- */ +//@cindex mkMutCons + static StgClosure * mkMutCons(StgClosure *ptr, generation *gen) { @@ -1001,7 +1266,7 @@ 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; } @@ -1030,7 +1295,7 @@ mkMutCons(StgClosure *ptr, generation *gen) didn't manage to evacuate this object into evac_gen. -------------------------------------------------------------------------- */ - +//@cindex evacuate static StgClosure * evacuate(StgClosure *q) @@ -1041,7 +1306,7 @@ evacuate(StgClosure *q) const StgInfoTable *info; loop: - if (!LOOKS_LIKE_STATIC(q)) { + if (HEAP_ALLOCED(q)) { bd = Bdescr((P_)q); if (bd->gen->no > N) { /* Can't evacuate this object, because it's in a generation @@ -1057,54 +1322,73 @@ loop: } step = bd->step->to; } +#ifdef DEBUG + else step = NULL; /* make sure copy() will crash if HEAP_ALLOCED is wrong */ +#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)))); - info = get_itbl(q); + /* + if (info->type==RBH) { + info = REVERT_INFOPTR(info); + IF_DEBUG(gc, + belch("@_ Trying to evacuate an RBH %p (%s); reverting to IP %p (%s)", + q, info_type(q), info, info_type_by_ip(info))); + } + */ + switch (info -> type) { case BCO: - to = copy(q,bco_sizeW(stgCast(StgBCO*,q)),step); - upd_evacuee(q,to); - return to; + { + nat size = bco_sizeW((StgBCO*)q); + + if (size >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) { + evacuate_large((P_)q, rtsFalse); + to = q; + } else { + /* just copy the block */ + to = copy(q,size,step); + } + return to; + } case MUT_VAR: + ASSERT(q->header.info != &MUT_CONS_info); case MVAR: to = copy(q,sizeW_fromITBL(info),step); - upd_evacuee(q,to); - evacuate_mutable((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); + recordMutable((StgMutClosure *)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; + return copy(q,sizeofW(StgHeader)+1,step); 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: +#ifdef NO_PROMOTE_THUNKS + if (bd->gen->no == 0 && + bd->step->no != 0 && + bd->step->no == bd->gen->n_steps-1) { + step = bd->step; + } +#endif + return copy(q,sizeofW(StgHeader)+2,step); + + case FUN_1_1: + case FUN_0_2: + case FUN_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); - return to; + return copy(q,sizeofW(StgHeader)+2,step); case FUN: case THUNK: @@ -1115,20 +1399,18 @@ loop: case CAF_ENTERED: case WEAK: case FOREIGN: - to = copy(q,sizeW_fromITBL(info),step); - upd_evacuee(q,to); - return to; + case STABLE_NAME: + return copy(q,sizeW_fromITBL(info),step); case CAF_BLACKHOLE: + case SE_CAF_BLACKHOLE: + case SE_BLACKHOLE: case BLACKHOLE: - to = copyPart(q,BLACKHOLE_sizeW(),sizeofW(StgHeader),step); - upd_evacuee(q,to); - return to; + return copyPart(q,BLACKHOLE_sizeW(),sizeofW(StgHeader),step); case BLACKHOLE_BQ: to = copy(q,BLACKHOLE_sizeW(),step); - upd_evacuee(q,to); - evacuate_mutable((StgMutClosure *)to); + recordMutable((StgMutClosure *)to); return to; case THUNK_SELECTOR: @@ -1147,11 +1429,11 @@ loop: case CONSTR_0_2: case CONSTR_STATIC: { - StgNat32 offset = info->layout.selector_offset; + StgWord32 offset = info->layout.selector_offset; /* check that the size is in range */ ASSERT(offset < - (StgNat32)(selectee_info->layout.payload.ptrs + + (StgWord32)(selectee_info->layout.payload.ptrs + selectee_info->layout.payload.nptrs)); /* perform the selection! */ @@ -1161,7 +1443,7 @@ loop: * with the evacuation, just update the source address with * a pointer to the (evacuated) constructor field. */ - if (IS_USER_PTR(q)) { + if (HEAP_ALLOCED(q)) { bdescr *bd = Bdescr((P_)q); if (bd->evacuated) { if (bd->gen->no < evac_gen) { @@ -1183,17 +1465,18 @@ loop: case IND_PERM: case IND_OLDGEN: case IND_OLDGEN_PERM: - selectee = stgCast(StgInd *,selectee)->indirectee; + selectee = ((StgInd *)selectee)->indirectee; goto selector_loop; case CAF_ENTERED: - selectee = stgCast(StgCAF *,selectee)->value; + selectee = ((StgCAF *)selectee)->value; goto selector_loop; case EVACUATED: - selectee = stgCast(StgEvacuated*,selectee)->evacuee; + selectee = ((StgEvacuated *)selectee)->evacuee; goto selector_loop; + case AP_UPD: case THUNK: case THUNK_1_0: case THUNK_0_1: @@ -1205,18 +1488,19 @@ loop: /* aargh - do recursively???? */ case CAF_UNENTERED: case CAF_BLACKHOLE: + case SE_CAF_BLACKHOLE: + case SE_BLACKHOLE: case BLACKHOLE: case BLACKHOLE_BQ: /* not evaluated yet */ break; default: - barf("evacuate: THUNK_SELECTOR: strange selectee"); + barf("evacuate: THUNK_SELECTOR: strange selectee %d", + (int)(selectee_info->type)); } } - to = copy(q,THUNK_SELECTOR_sizeW(),step); - upd_evacuee(q,to); - return to; + return copy(q,THUNK_SELECTOR_sizeW(),step); case IND: case IND_OLDGEN: @@ -1224,30 +1508,35 @@ loop: q = ((StgInd*)q)->indirectee; goto loop; - /* ToDo: optimise STATIC_LINK for known cases. - - FUN_STATIC : payload[0] - - THUNK_STATIC : payload[1] - - IND_STATIC : payload[1] - */ case THUNK_STATIC: + if (info->srt_len > 0 && major_gc && + THUNK_STATIC_LINK((StgClosure *)q) == NULL) { + THUNK_STATIC_LINK((StgClosure *)q) = static_objects; + static_objects = (StgClosure *)q; + } + return q; + case FUN_STATIC: - if (info->srt_len == 0) { /* small optimisation */ - return q; + if (info->srt_len > 0 && major_gc && + FUN_STATIC_LINK((StgClosure *)q) == NULL) { + FUN_STATIC_LINK((StgClosure *)q) = static_objects; + static_objects = (StgClosure *)q; } - /* fall through */ - case CONSTR_STATIC: + return q; + case IND_STATIC: - /* don't want to evacuate these, but we do want to follow pointers - * from SRTs - see scavenge_static. - */ + if (major_gc && IND_STATIC_LINK((StgClosure *)q) == NULL) { + IND_STATIC_LINK((StgClosure *)q) = static_objects; + static_objects = (StgClosure *)q; + } + return q; - /* put the object on the static list, if necessary. - */ + case CONSTR_STATIC: if (major_gc && STATIC_LINK(info,(StgClosure *)q) == NULL) { STATIC_LINK(info,(StgClosure *)q) = static_objects; static_objects = (StgClosure *)q; } - /* fall through */ + return q; case CONSTR_INTLIKE: case CONSTR_CHARLIKE: @@ -1268,15 +1557,26 @@ loop: case CATCH_FRAME: case SEQ_FRAME: /* shouldn't see these */ - barf("evacuate: stack frame\n"); + barf("evacuate: stack frame at %p\n", q); case AP_UPD: 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)),step); - upd_evacuee(q,to); - return to; + /* PAPs and AP_UPDs are special - the payload is a copy of a chunk + * of stack, tagging and all. + * + * They can be larger than a block in size. Both are only + * allocated via allocate(), so they should be chained on to the + * large_object list. + */ + { + nat size = pap_sizeW((StgPAP*)q); + if (size >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) { + evacuate_large((P_)q, rtsFalse); + return q; + } else { + return copy(q,size,step); + } + } case EVACUATED: /* Already evacuated, just return the forwarding address. @@ -1289,33 +1589,30 @@ loop: if (evac_gen > 0) { /* optimisation */ StgClosure *p = ((StgEvacuated*)q)->evacuee; if (Bdescr((P_)p)->gen->no < evac_gen) { - /* fprintf(stderr,"evac failed!\n");*/ + IF_DEBUG(gc, belch("@@ evacuate: evac of EVACUATED node %p failed!", p)); failed_to_evac = rtsTrue; TICK_GC_FAILED_PROMOTION(); - } + } } return ((StgEvacuated*)q)->evacuee; - case MUT_ARR_WORDS: case ARR_WORDS: { - nat size = arr_words_sizeW(stgCast(StgArrWords*,q)); + nat size = arr_words_sizeW((StgArrWords *)q); if (size >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) { evacuate_large((P_)q, rtsFalse); return q; } else { /* just copy the block */ - to = copy(q,size,step); - upd_evacuee(q,to); - return to; + return copy(q,size,step); } } case MUT_ARR_PTRS: case MUT_ARR_PTRS_FROZEN: { - nat size = mut_arr_ptrs_sizeW(stgCast(StgMutArrPtrs*,q)); + nat size = mut_arr_ptrs_sizeW((StgMutArrPtrs *)q); if (size >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) { evacuate_large((P_)q, info->type == MUT_ARR_PTRS); @@ -1323,9 +1620,8 @@ loop: } else { /* just copy the block */ to = copy(q,size,step); - upd_evacuee(q,to); if (info->type == MUT_ARR_PTRS) { - evacuate_mutable((StgMutClosure *)to); + recordMutable((StgMutClosure *)to); } } return to; @@ -1333,10 +1629,17 @@ loop: case TSO: { - StgTSO *tso = stgCast(StgTSO *,q); + StgTSO *tso = (StgTSO *)q; nat size = tso_sizeW(tso); int diff; + /* Deal with redirected TSOs (a TSO that's had its stack enlarged). + */ + if (tso->what_next == ThreadRelocated) { + q = (StgClosure *)tso->link; + goto loop; + } + /* Large TSOs don't get moved, so no relocation is required. */ if (size >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) { @@ -1354,23 +1657,55 @@ loop: /* relocate the stack pointers... */ new_tso->su = (StgUpdateFrame *) ((StgPtr)new_tso->su + diff); new_tso->sp = (StgPtr)new_tso->sp + diff; - new_tso->splim = (StgPtr)new_tso->splim + diff; relocate_TSO(tso, new_tso); - upd_evacuee(q,(StgClosure *)new_tso); - evacuate_mutable((StgMutClosure *)new_tso); + recordMutable((StgMutClosure *)new_tso); return (StgClosure *)new_tso; } } +#if defined(PAR) + case RBH: // cf. BLACKHOLE_BQ + { + //StgInfoTable *rip = get_closure_info(q, &size, &ptrs, &nonptrs, &vhs, str); + to = copy(q,BLACKHOLE_sizeW(),step); + //ToDo: derive size etc from reverted IP + //to = copy(q,size,step); + recordMutable((StgMutClosure *)to); + IF_DEBUG(gc, + belch("@@ evacuate: RBH %p (%s) to %p (%s)", + q, info_type(q), to, info_type(to))); + return to; + } + case BLOCKED_FETCH: + ASSERT(sizeofW(StgBlockedFetch) >= MIN_NONUPD_SIZE); + to = copy(q,sizeofW(StgBlockedFetch),step); + IF_DEBUG(gc, + belch("@@ evacuate: %p (%s) to %p (%s)", + q, info_type(q), to, info_type(to))); + return to; + case FETCH_ME: - fprintf(stderr,"evacuate: unimplemented/strange closure type\n"); - return q; + ASSERT(sizeofW(StgBlockedFetch) >= MIN_UPD_SIZE); + to = copy(q,sizeofW(StgFetchMe),step); + IF_DEBUG(gc, + belch("@@ evacuate: %p (%s) to %p (%s)", + q, info_type(q), to, info_type(to))); + return to; + + case FETCH_ME_BQ: + ASSERT(sizeofW(StgBlockedFetch) >= MIN_UPD_SIZE); + to = copy(q,sizeofW(StgFetchMeBlockingQueue),step); + IF_DEBUG(gc, + belch("@@ evacuate: %p (%s) to %p (%s)", + q, info_type(q), to, info_type(to))); + return to; +#endif default: - barf("evacuate: strange closure type"); + barf("evacuate: strange closure type %d", (int)(info->type)); } barf("evacuate"); @@ -1380,6 +1715,7 @@ loop: relocate_TSO is called just after a TSO has been copied from src to dest. It adjusts the update frame list for the new location. -------------------------------------------------------------------------- */ +//@cindex relocate_TSO StgTSO * relocate_TSO(StgTSO *src, StgTSO *dest) @@ -1420,7 +1756,7 @@ relocate_TSO(StgTSO *src, StgTSO *dest) break; default: - barf("relocate_TSO"); + barf("relocate_TSO %d", (int)(get_itbl(su)->type)); } break; } @@ -1428,6 +1764,11 @@ relocate_TSO(StgTSO *src, StgTSO *dest) return dest; } +//@node Scavenging, Reverting CAFs, Evacuation +//@subsection Scavenging + +//@cindex scavenge_srt + static inline void scavenge_srt(const StgInfoTable *info) { @@ -1437,14 +1778,58 @@ scavenge_srt(const StgInfoTable *info) * srt field in the info table. That's ok, because we'll * never dereference it. */ - srt = stgCast(StgClosure **,info->srt); + srt = (StgClosure **)(info->srt); srt_end = srt + info->srt_len; for (; srt < srt_end; srt++) { - evacuate(*srt); + /* Special-case to handle references to closures hiding out in DLLs, since + double indirections required to get at those. The code generator knows + which is which when generating the SRT, so it stores the (indirect) + reference to the DLL closure in the table by first adding one to it. + We check for this here, and undo the addition before evacuating it. + + If the SRT entry hasn't got bit 0 set, the SRT entry points to a + closure that's fixed at link-time, and no extra magic is required. + */ +#ifdef ENABLE_WIN32_DLL_SUPPORT + if ( (unsigned long)(*srt) & 0x1 ) { + evacuate(*stgCast(StgClosure**,(stgCast(unsigned long, *srt) & ~0x1))); + } else { + evacuate(*srt); + } +#else + evacuate(*srt); +#endif } } /* ----------------------------------------------------------------------------- + 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 +#if defined(PAR) + || 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])); +} + +/* ----------------------------------------------------------------------------- Scavenge a given step until there are no more objects in this step to scavenge. @@ -1456,7 +1841,7 @@ scavenge_srt(const StgInfoTable *info) scavenging a mutable object where early promotion isn't such a good idea. -------------------------------------------------------------------------- */ - +//@cindex scavenge static void scavenge(step *step) @@ -1490,11 +1875,16 @@ scavenge(step *step) || IS_HUGS_CONSTR_INFO(GET_INFO((StgClosure *)p)))); info = get_itbl((StgClosure *)p); + /* + if (info->type==RBH) + info = REVERT_INFOPTR(info); + */ + switch (info -> type) { case BCO: { - StgBCO* bco = stgCast(StgBCO*,p); + StgBCO* bco = (StgBCO *)p; nat i; for (i = 0; i < bco->n_ptrs; i++) { bcoConstCPtr(bco,i) = evacuate(bcoConstCPtr(bco,i)); @@ -1575,10 +1965,6 @@ scavenge(step *step) case WEAK: case FOREIGN: case STABLE_NAME: - case IND_PERM: - case IND_OLDGEN_PERM: - case CAF_UNENTERED: - case CAF_ENTERED: { StgPtr end; @@ -1586,7 +1972,53 @@ scavenge(step *step) for (p = (P_)((StgClosure *)p)->payload; p < end; p++) { (StgClosure *)*p = evacuate((StgClosure *)*p); } - p += info->layout.payload.nptrs; + p += info->layout.payload.nptrs; + break; + } + + case IND_PERM: + if (step->gen->no != 0) { + SET_INFO(((StgClosure *)p), &IND_OLDGEN_PERM_info); + } + /* fall through */ + case IND_OLDGEN_PERM: + ((StgIndOldGen *)p)->indirectee = + evacuate(((StgIndOldGen *)p)->indirectee); + if (failed_to_evac) { + failed_to_evac = rtsFalse; + recordOldToNewPtrs((StgMutClosure *)p); + } + p += sizeofW(StgIndOldGen); + break; + + case CAF_UNENTERED: + { + StgCAF *caf = (StgCAF *)p; + + caf->body = evacuate(caf->body); + if (failed_to_evac) { + failed_to_evac = rtsFalse; + recordOldToNewPtrs((StgMutClosure *)p); + } else { + caf->mut_link = NULL; + } + p += sizeofW(StgCAF); + break; + } + + case CAF_ENTERED: + { + StgCAF *caf = (StgCAF *)p; + + caf->body = evacuate(caf->body); + caf->value = evacuate(caf->value); + if (failed_to_evac) { + failed_to_evac = rtsFalse; + recordOldToNewPtrs((StgMutClosure *)p); + } else { + caf->mut_link = NULL; + } + p += sizeofW(StgCAF); break; } @@ -1601,6 +2033,8 @@ scavenge(step *step) break; case CAF_BLACKHOLE: + case SE_CAF_BLACKHOLE: + case SE_BLACKHOLE: case BLACKHOLE: p += BLACKHOLE_sizeW(); break; @@ -1612,7 +2046,7 @@ scavenge(step *step) evacuate((StgClosure *)bh->blocking_queue); if (failed_to_evac) { failed_to_evac = rtsFalse; - evacuate_mutable((StgMutClosure *)bh); + recordMutable((StgMutClosure *)bh); } p += BLACKHOLE_sizeW(); break; @@ -1659,7 +2093,7 @@ scavenge(step *step) * 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); @@ -1668,9 +2102,8 @@ scavenge(step *step) } case ARR_WORDS: - case MUT_ARR_WORDS: /* nothing to follow */ - p += arr_words_sizeW(stgCast(StgArrWords*,p)); + p += arr_words_sizeW((StgArrWords *)p); break; case MUT_ARR_PTRS: @@ -1698,7 +2131,7 @@ scavenge(step *step) } if (failed_to_evac) { /* we can do this easier... */ - evacuate_mutable((StgMutClosure *)start); + recordMutable((StgMutClosure *)start); failed_to_evac = rtsFalse; } break; @@ -1706,26 +2139,87 @@ scavenge(step *step) case TSO: { - StgTSO *tso; - - tso = (StgTSO *)p; + StgTSO *tso = (StgTSO *)p; evac_gen = 0; - /* chase the link field for any TSOs on the same queue */ - (StgClosure *)tso->link = evacuate((StgClosure *)tso->link); - /* scavenge this thread's stack */ - scavenge_stack(tso->sp, &(tso->stack[tso->stack_size])); + scavengeTSO(tso); evac_gen = saved_evac_gen; p += tso_sizeW(tso); break; } +#if defined(PAR) + case RBH: // cf. BLACKHOLE_BQ + { + // nat size, ptrs, nonptrs, vhs; + // char str[80]; + // StgInfoTable *rip = get_closure_info(p, &size, &ptrs, &nonptrs, &vhs, str); + StgRBH *rbh = (StgRBH *)p; + (StgClosure *)rbh->blocking_queue = + evacuate((StgClosure *)rbh->blocking_queue); + if (failed_to_evac) { + failed_to_evac = rtsFalse; + recordMutable((StgMutClosure *)rbh); + } + IF_DEBUG(gc, + belch("@@ scavenge: RBH %p (%s) (new blocking_queue link=%p)", + p, info_type(p), (StgClosure *)rbh->blocking_queue)); + // ToDo: use size of reverted closure here! + p += BLACKHOLE_sizeW(); + break; + } + case BLOCKED_FETCH: + { + StgBlockedFetch *bf = (StgBlockedFetch *)p; + /* follow the pointer to the node which is being demanded */ + (StgClosure *)bf->node = + evacuate((StgClosure *)bf->node); + /* follow the link to the rest of the blocking queue */ + (StgClosure *)bf->link = + evacuate((StgClosure *)bf->link); + if (failed_to_evac) { + failed_to_evac = rtsFalse; + recordMutable((StgMutClosure *)bf); + } + IF_DEBUG(gc, + belch("@@ scavenge: %p (%s); node is now %p; exciting, isn't it", + bf, info_type((StgClosure *)bf), + bf->node, info_type(bf->node))); + p += sizeofW(StgBlockedFetch); + break; + } + case FETCH_ME: + IF_DEBUG(gc, + belch("@@ scavenge: HWL claims nothing to do for %p (%s)", + p, info_type((StgClosure *)p))); + p += sizeofW(StgFetchMe); + break; // nothing to do in this case + + case FETCH_ME_BQ: // cf. BLACKHOLE_BQ + { + StgFetchMeBlockingQueue *fmbq = (StgFetchMeBlockingQueue *)p; + (StgClosure *)fmbq->blocking_queue = + evacuate((StgClosure *)fmbq->blocking_queue); + if (failed_to_evac) { + failed_to_evac = rtsFalse; + recordMutable((StgMutClosure *)fmbq); + } + IF_DEBUG(gc, + belch("@@ scavenge: %p (%s) exciting, isn't it", + p, info_type((StgClosure *)p))); + p += sizeofW(StgFetchMeBlockingQueue); + break; + } +#endif + case EVACUATED: - barf("scavenge: unimplemented/strange closure type\n"); + barf("scavenge: unimplemented/strange closure type %d @ %p", + info->type, p); default: - barf("scavenge"); + barf("scavenge: unimplemented/strange closure type %d @ %p", + info->type, p); } /* If we didn't manage to promote all the objects pointed to by @@ -1749,16 +2243,23 @@ scavenge(step *step) because they contain old-to-new generation pointers. Only certain objects can have this property. -------------------------------------------------------------------------- */ +//@cindex scavenge_one + static rtsBool -scavenge_one(StgPtr p) +scavenge_one(StgClosure *p) { - StgInfoTable *info; + const 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); + + /* ngoq moHqu'! + if (info->type==RBH) + info = REVERT_INFOPTR(info); // if it's an RBH, look at the orig closure + */ switch (info -> type) { @@ -1785,18 +2286,19 @@ scavenge_one(StgPtr p) case IND_PERM: case IND_OLDGEN_PERM: 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 SE_CAF_BLACKHOLE: + case SE_BLACKHOLE: case BLACKHOLE: break; @@ -1804,7 +2306,7 @@ scavenge_one(StgPtr p) { StgSelector *s = (StgSelector *)p; s->selectee = evacuate(s->selectee); - break; + break; } case AP_UPD: /* same as PAPs */ @@ -1813,7 +2315,7 @@ 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); @@ -1829,7 +2331,7 @@ scavenge_one(StgPtr p) break; default: - barf("scavenge_one: strange object"); + barf("scavenge_one: strange object %d", (int)(info->type)); } no_luck = failed_to_evac; @@ -1845,28 +2347,168 @@ scavenge_one(StgPtr p) generations older than the one being collected) as roots. We also remove non-mutable objects from the mutable list at this point. -------------------------------------------------------------------------- */ +//@cindex scavenge_mut_once_list -static StgMutClosure * -scavenge_mutable_list(StgMutClosure *p, nat gen) +static void +scavenge_mut_once_list(generation *gen) { - StgInfoTable *info; - StgMutClosure *start; - StgMutClosure **prev; + const StgInfoTable *info; + StgMutClosure *p, *next, *new_list; - evac_gen = 0; + p = gen->mut_once_list; + new_list = END_MUT_LIST; + next = p->mut_link; + + evac_gen = gen->no; + failed_to_evac = rtsFalse; + + 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)))); + + info = get_itbl(p); + /* + if (info->type==RBH) + info = REVERT_INFOPTR(info); // if it's an RBH, look at the orig closure + */ + 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); + +#ifdef DEBUG + if (RtsFlags.DebugFlags.gc) + /* Debugging code to print out the size of the thing we just + * promoted + */ + { + StgPtr start = gen->steps[0].scan; + bdescr *start_bd = gen->steps[0].scan_bd; + nat size = 0; + scavenge(&gen->steps[0]); + if (start_bd != gen->steps[0].scan_bd) { + size += (P_)BLOCK_ROUND_UP(start) - start; + start_bd = start_bd->link; + while (start_bd != gen->steps[0].scan_bd) { + size += BLOCK_SIZE_W; + start_bd = start_bd->link; + } + size += gen->steps[0].scan - + (P_)BLOCK_ROUND_DOWN(gen->steps[0].scan); + } else { + size = gen->steps[0].scan - start; + } + fprintf(stderr,"evac IND_OLDGEN: %d bytes\n", size * sizeof(W_)); + } +#endif + + /* 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; + p->mut_link = new_list; + new_list = p; + } else { + /* 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 put the + * MUT_CONS back on the list. + */ + p->mut_link = new_list; + new_list = p; + } + continue; + + case CAF_ENTERED: + { + StgCAF *caf = (StgCAF *)p; + caf->body = evacuate(caf->body); + caf->value = evacuate(caf->value); + if (failed_to_evac) { + failed_to_evac = rtsFalse; + p->mut_link = new_list; + new_list = p; + } else { + p->mut_link = NULL; + } + } + continue; + + case CAF_UNENTERED: + { + StgCAF *caf = (StgCAF *)p; + caf->body = evacuate(caf->body); + if (failed_to_evac) { + failed_to_evac = rtsFalse; + p->mut_link = new_list; + new_list = p; + } else { + p->mut_link = NULL; + } + } + continue; + + default: + /* shouldn't have anything else on the mutables list */ + barf("scavenge_mut_once_list: strange object? %d", (int)(info->type)); + } + } - prev = &start; - start = p; + gen->mut_once_list = new_list; +} + +//@cindex scavenge_mutable_list + +static void +scavenge_mutable_list(generation *gen) +{ + const StgInfoTable *info; + StgMutClosure *p, *next; + p = gen->saved_mut_list; + next = p->mut_link; + + evac_gen = 0; failed_to_evac = rtsFalse; - for (; p != END_MUT_LIST; p = *prev) { + 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)))); info = get_itbl(p); + /* + if (info->type==RBH) + info = REVERT_INFOPTR(info); // if it's an RBH, look at the orig closure + */ switch(info->type) { case MUT_ARR_PTRS_FROZEN: @@ -1877,7 +2519,7 @@ scavenge_mutable_list(StgMutClosure *p, nat gen) StgPtr end, q; end = (P_)p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p); - evac_gen = gen; + evac_gen = gen->no; for (q = (P_)((StgMutArrPtrs *)p)->payload; q < end; q++) { (StgClosure *)*q = evacuate((StgClosure *)*q); } @@ -1885,16 +2527,16 @@ scavenge_mutable_list(StgMutClosure *p, nat gen) if (failed_to_evac) { failed_to_evac = rtsFalse; - prev = &p->mut_link; - } else { - *prev = p->mut_link; - } + p->mut_link = gen->mut_list; + gen->mut_list = p; + } continue; } case MUT_ARR_PTRS: /* follow everything */ - prev = &p->mut_link; + p->mut_link = gen->mut_list; + gen->mut_list = p; { StgPtr end, q; @@ -1910,21 +2552,10 @@ 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); + p->mut_link = gen->mut_list; + gen->mut_list = p; continue; case MVAR: @@ -1933,77 +2564,121 @@ scavenge_mutable_list(StgMutClosure *p, nat gen) (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; + p->mut_link = gen->mut_list; + gen->mut_list = p; continue; } case TSO: - /* follow ptrs and remove this from the mutable list */ { StgTSO *tso = (StgTSO *)p; - /* Don't bother scavenging if this thread is dead - */ - if (!(tso->whatNext == ThreadComplete || - tso->whatNext == ThreadKilled)) { - /* Don't need to chase the link field for any TSOs on the - * same queue. Just scavenge this thread's stack - */ - scavenge_stack(tso->sp, &(tso->stack[tso->stack_size])); - } + scavengeTSO(tso); /* Don't take this TSO off the mutable list - it might still * point to some younger objects (because we set evac_gen to 0 * above). */ - prev = &tso->mut_link; + tso->mut_link = gen->mut_list; + gen->mut_list = (StgMutClosure *)tso; continue; } + case BLACKHOLE_BQ: + { + StgBlockingQueue *bh = (StgBlockingQueue *)p; + (StgClosure *)bh->blocking_queue = + evacuate((StgClosure *)bh->blocking_queue); + p->mut_link = gen->mut_list; + gen->mut_list = p; + continue; + } + + /* Happens if a BLACKHOLE_BQ in the old generation is updated: + */ 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; + evac_gen = gen->no; ((StgIndOldGen *)p)->indirectee = evacuate(((StgIndOldGen *)p)->indirectee); evac_gen = 0; if (failed_to_evac) { failed_to_evac = rtsFalse; - prev = &p->mut_link; + p->mut_link = gen->mut_once_list; + gen->mut_once_list = p; } 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 BLACKHOLE_BQ: + +#if defined(PAR) + // HWL: check whether all of these are necessary + + case RBH: // cf. BLACKHOLE_BQ { - StgBlockingQueue *bh = (StgBlockingQueue *)p; - (StgClosure *)bh->blocking_queue = - evacuate((StgClosure *)bh->blocking_queue); - prev = &p->mut_link; + // nat size, ptrs, nonptrs, vhs; + // char str[80]; + // StgInfoTable *rip = get_closure_info(p, &size, &ptrs, &nonptrs, &vhs, str); + StgRBH *rbh = (StgRBH *)p; + (StgClosure *)rbh->blocking_queue = + evacuate((StgClosure *)rbh->blocking_queue); + if (failed_to_evac) { + failed_to_evac = rtsFalse; + recordMutable((StgMutClosure *)rbh); + } + // ToDo: use size of reverted closure here! + p += BLACKHOLE_sizeW(); + break; + } + + case BLOCKED_FETCH: + { + StgBlockedFetch *bf = (StgBlockedFetch *)p; + /* follow the pointer to the node which is being demanded */ + (StgClosure *)bf->node = + evacuate((StgClosure *)bf->node); + /* follow the link to the rest of the blocking queue */ + (StgClosure *)bf->link = + evacuate((StgClosure *)bf->link); + if (failed_to_evac) { + failed_to_evac = rtsFalse; + recordMutable((StgMutClosure *)bf); + } + p += sizeofW(StgBlockedFetch); + break; + } + + case FETCH_ME: + p += sizeofW(StgFetchMe); + break; // nothing to do in this case + + case FETCH_ME_BQ: // cf. BLACKHOLE_BQ + { + StgFetchMeBlockingQueue *fmbq = (StgFetchMeBlockingQueue *)p; + (StgClosure *)fmbq->blocking_queue = + evacuate((StgClosure *)fmbq->blocking_queue); + if (failed_to_evac) { + failed_to_evac = rtsFalse; + recordMutable((StgMutClosure *)fmbq); + } + p += sizeofW(StgFetchMeBlockingQueue); break; } +#endif default: /* shouldn't have anything else on the mutables list */ - barf("scavenge_mutable_object: non-mutable object?"); + barf("scavenge_mutable_list: strange object? %d", (int)(info->type)); } } - return start; } +//@cindex scavenge_static + static void scavenge_static(void) { @@ -2019,7 +2694,10 @@ scavenge_static(void) while (p != END_OF_STATIC_LIST) { 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)))); @@ -2046,8 +2724,8 @@ scavenge_static(void) if (failed_to_evac) { failed_to_evac = rtsFalse; scavenged_static_objects = STATIC_LINK(info,p); - ((StgMutClosure *)ind)->mut_link = oldest_gen->mut_list; - oldest_gen->mut_list = (StgMutClosure *)ind; + ((StgMutClosure *)ind)->mut_link = oldest_gen->mut_once_list; + oldest_gen->mut_once_list = (StgMutClosure *)ind; } break; } @@ -2070,12 +2748,12 @@ scavenge_static(void) } default: - barf("scavenge_static"); + barf("scavenge_static: strange closure %d", (int)(info->type)); } ASSERT(failed_to_evac == rtsFalse); - /* get the next static object from the list. Remeber, there might + /* get the next static object from the list. Remember, there might * be more stuff on this list now that we've done some evacuating! * (static_objects is a global) */ @@ -2088,13 +2766,16 @@ scavenge_static(void) objects pointed to by it. We can use the same code for walking PAPs, since these are just sections of copied stack. -------------------------------------------------------------------------- */ +//@cindex scavenge_stack static void scavenge_stack(StgPtr p, StgPtr stack_end) { StgPtr q; const StgInfoTable* info; - StgNat32 bitmap; + StgWord32 bitmap; + + //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 @@ -2103,24 +2784,22 @@ scavenge_stack(StgPtr p, StgPtr stack_end) */ while (p < stack_end) { - q = *stgCast(StgPtr*,p); + q = *(P_ *)p; /* If we've got a tag, skip over that many words on the stack */ - if (IS_ARG_TAG(stgCast(StgWord,q))) { + 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)) { - + if (! LOOKS_LIKE_GHC_INFO(q) ) { #ifdef DEBUG - if (LOOKS_LIKE_STATIC(q)) { /* Is it a static closure? */ - ASSERT(closure_STATIC(stgCast(StgClosure*,q))); - } - /* otherwise, must be a pointer into the allocation space. - */ + 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); @@ -2133,21 +2812,31 @@ scavenge_stack(StgPtr p, StgPtr stack_end) * record. All activation records have 'bitmap' style layout * info. */ - info = get_itbl(stgCast(StgClosure*,p)); + info = get_itbl((StgClosure *)p); switch (info->type) { /* Dynamic bitmap: the mask is stored on the stack */ case RET_DYN: - bitmap = stgCast(StgRetDyn*,p)->liveness; - p = &payloadWord(stgCast(StgRetDyn*,p),0); + 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: - p++; + { +#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, @@ -2157,7 +2846,7 @@ scavenge_stack(StgPtr p, StgPtr stack_end) { StgUpdateFrame *frame = (StgUpdateFrame *)p; StgClosure *to; - StgClosureType type = get_itbl(frame->updatee)->type; + nat type = get_itbl(frame->updatee)->type; p += sizeofW(StgUpdateFrame); if (type == EVACUATED) { @@ -2172,36 +2861,45 @@ scavenge_stack(StgPtr p, StgPtr stack_end) } continue; } - step = bd->step->to; + + /* Don't promote blackholes */ + step = bd->step; + if (!(step->gen->no == 0 && + step->no != 0 && + step->no == step->gen->n_steps-1)) { + step = 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; - evacuate_mutable((StgMutClosure *)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"); } } } /* small bitmap (< 32 entries, or 64 on a 64-bit machine) */ - case RET_BCO: - case RET_SMALL: - case RET_VEC_SMALL: 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) { @@ -2249,7 +2947,7 @@ scavenge_stack(StgPtr p, StgPtr stack_end) } default: - barf("scavenge_stack: weird activation record found on stack.\n"); + barf("scavenge_stack: weird activation record found on stack: %d", (int)(info->type)); } } } @@ -2262,6 +2960,7 @@ scavenge_stack(StgPtr p, StgPtr stack_end) objects are (repeatedly) mutable, so most of the time evac_gen will be zero. --------------------------------------------------------------------------- */ +//@cindex scavenge_large static void scavenge_large(step *step) @@ -2285,14 +2984,13 @@ scavenge_large(step *step) dbl_link_onto(bd, &step->scavenged_large_objects); p = bd->start; - info = get_itbl(stgCast(StgClosure*,p)); + info = get_itbl((StgClosure *)p); switch (info->type) { /* only certain objects can be "large"... */ case ARR_WORDS: - case MUT_ARR_WORDS: /* nothing to follow */ continue; @@ -2320,14 +3018,14 @@ scavenge_large(step *step) } evac_gen = 0; if (failed_to_evac) { - evacuate_mutable((StgMutClosure *)start); + recordMutable((StgMutClosure *)start); } continue; } case BCO: { - StgBCO* bco = stgCast(StgBCO*,p); + StgBCO* bco = (StgBCO *)p; nat i; evac_gen = saved_evac_gen; for (i = 0; i < bco->n_ptrs; i++) { @@ -2338,25 +3036,31 @@ scavenge_large(step *step) } case TSO: + scavengeTSO((StgTSO *)p); + continue; + + case AP_UPD: + case PAP: { - StgTSO *tso; + StgPAP* pap = (StgPAP *)p; - tso = (StgTSO *)p; - /* chase the link field for any TSOs on the same queue */ - (StgClosure *)tso->link = evacuate((StgClosure *)tso->link); - /* scavenge this thread's stack */ - scavenge_stack(tso->sp, &(tso->stack[tso->stack_size])); + evac_gen = saved_evac_gen; /* not really mutable */ + pap->fun = evacuate(pap->fun); + scavenge_stack((P_)pap->payload, (P_)pap->payload + pap->n_args); + evac_gen = 0; continue; } default: - barf("scavenge_large: unknown/strange object"); + barf("scavenge_large: unknown/strange object %d", (int)(info->type)); } } } +//@cindex zero_static_object_list + static void -zeroStaticObjectList(StgClosure* first_static) +zero_static_object_list(StgClosure* first_static) { StgClosure* p; StgClosure* link; @@ -2377,8 +3081,10 @@ zeroStaticObjectList(StgClosure* first_static) * It doesn't do any harm to zero all the mutable link fields on the * mutable list. */ +//@cindex zero_mutable_list + static void -zeroMutableList(StgMutClosure *first) +zero_mutable_list( StgMutClosure *first ) { StgMutClosure *next, *c; @@ -2388,52 +3094,43 @@ zeroMutableList(StgMutClosure *first) } } +//@node Reverting CAFs, Sanity code for CAF garbage collection, Scavenging +//@subsection Reverting CAFs + /* ----------------------------------------------------------------------------- Reverting CAFs -------------------------------------------------------------------------- */ +//@cindex RevertCAFs void RevertCAFs(void) { - while (enteredCAFs != END_CAF_LIST) { - StgCAF* caf = enteredCAFs; - - enteredCAFs = caf->link; - ASSERT(get_itbl(caf)->type == CAF_ENTERED); - SET_INFO(caf,&CAF_UNENTERED_info); - caf->value = stgCast(StgClosure*,0xdeadbeef); - caf->link = stgCast(StgCAF*,0xdeadbeef); - } +#ifdef INTERPRETER + StgInt i; + + /* Deal with CAFs created by compiled code. */ + for (i = 0; i < usedECafTable; i++) { + SET_INFO( (StgInd*)(ecafTable[i].closure), ecafTable[i].origItbl ); + ((StgInd*)(ecafTable[i].closure))->indirectee = 0; + } + + /* Deal with CAFs created by the interpreter. */ + while (ecafList != END_ECAF_LIST) { + StgCAF* caf = ecafList; + ecafList = caf->link; + ASSERT(get_itbl(caf)->type == CAF_ENTERED); + SET_INFO(caf,&CAF_UNENTERED_info); + caf->value = (StgClosure *)0xdeadbeef; + caf->link = (StgCAF *)0xdeadbeef; + } + + /* Empty out both the table and the list. */ + clearECafTable(); + ecafList = END_ECAF_LIST; +#endif } -void revertDeadCAFs(void) -{ - StgCAF* caf = enteredCAFs; - enteredCAFs = END_CAF_LIST; - while (caf != END_CAF_LIST) { - StgCAF* next = caf->link; - - switch(GET_INFO(caf)->type) { - case EVACUATED: - { - /* This object has been evacuated, it must be live. */ - StgCAF* new = stgCast(StgCAF*,stgCast(StgEvacuated*,caf)->evacuee); - new->link = enteredCAFs; - enteredCAFs = new; - break; - } - case CAF_ENTERED: - { - SET_INFO(caf,&CAF_UNENTERED_info); - caf->value = stgCast(StgClosure*,0xdeadbeef); - caf->link = stgCast(StgCAF*,0xdeadbeef); - break; - } - default: - barf("revertDeadCAFs: enteredCAFs list corrupted"); - } - caf = next; - } -} +//@node Sanity code for CAF garbage collection, Lazy black holing, Reverting CAFs +//@subsection Sanity code for CAF garbage collection /* ----------------------------------------------------------------------------- Sanity code for CAF garbage collection. @@ -2448,6 +3145,8 @@ void revertDeadCAFs(void) -------------------------------------------------------------------------- */ #ifdef DEBUG +//@cindex gcCAFs + static void gcCAFs(void) { @@ -2485,6 +3184,9 @@ gcCAFs(void) } #endif +//@node Lazy black holing, Stack squeezing, Sanity code for CAF garbage collection +//@subsection Lazy black holing + /* ----------------------------------------------------------------------------- Lazy black holing. @@ -2492,6 +3194,7 @@ gcCAFs(void) some work, we have to run down the stack and black-hole all the closures referred to by update frames. -------------------------------------------------------------------------- */ +//@cindex threadLazyBlackHole static void threadLazyBlackHole(StgTSO *tso) @@ -2507,7 +3210,7 @@ threadLazyBlackHole(StgTSO *tso) switch (get_itbl(update_frame)->type) { case CATCH_FRAME: - update_frame = stgCast(StgCatchFrame*,update_frame)->link; + update_frame = ((StgCatchFrame *)update_frame)->link; break; case UPDATE_FRAME: @@ -2526,6 +3229,9 @@ threadLazyBlackHole(StgTSO *tso) if (bh->header.info != &BLACKHOLE_BQ_info && bh->header.info != &CAF_BLACKHOLE_info) { +#if (!defined(LAZY_BLACKHOLING)) && defined(DEBUG) + fprintf(stderr,"Unexpected lazy BHing required at 0x%04x\n",(int)bh); +#endif SET_INFO(bh,&BLACKHOLE_info); } @@ -2533,7 +3239,7 @@ threadLazyBlackHole(StgTSO *tso) break; case SEQ_FRAME: - update_frame = stgCast(StgSeqFrame*,update_frame)->link; + update_frame = ((StgSeqFrame *)update_frame)->link; break; case STOP_FRAME: @@ -2544,6 +3250,9 @@ threadLazyBlackHole(StgTSO *tso) } } +//@node Stack squeezing, Pausing a thread, Lazy black holing +//@subsection Stack squeezing + /* ----------------------------------------------------------------------------- * Stack squeezing * @@ -2551,6 +3260,7 @@ threadLazyBlackHole(StgTSO *tso) * lazy black holing here. * * -------------------------------------------------------------------------- */ +//@cindex threadSqueezeStack static void threadSqueezeStack(StgTSO *tso) @@ -2561,6 +3271,14 @@ threadSqueezeStack(StgTSO *tso) 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 + + top_frame = tso->su; +#endif bottom = &(tso->stack[tso->stack_size]); frame = tso->su; @@ -2580,11 +3298,36 @@ threadSqueezeStack(StgTSO *tso) */ next_frame = NULL; - while ((P_)frame < bottom - 1) { /* bottom - 1 is the STOP_FRAME */ + /* 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 == &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 == &BLACKHOLE_info) { break; @@ -2634,8 +3377,9 @@ threadSqueezeStack(StgTSO *tso) StgClosure *updatee_keep = prev_frame->updatee; StgClosure *updatee_bypass = frame->updatee; -#if 0 /* DEBUG */ - fprintf(stderr, "squeezing frame at %p\n", frame); +#if DEBUG + IF_DEBUG(gc, fprintf(stderr, "@@ squeezing frame at %p\n", frame)); + squeezes++; #endif /* Deal with blocking queues. If both updatees have blocked @@ -2648,7 +3392,12 @@ threadSqueezeStack(StgTSO *tso) * slower --SDM */ #if 0 /* do it properly... */ - if (GET_INFO(updatee_bypass) == BLACKHOLE_BQ_info) { +# if (!defined(LAZY_BLACKHOLING)) && defined(DEBUG) +# error Unimplemented lazy BH warning. (KSW 1999-01) +# endif + if (GET_INFO(updatee_bypass) == BLACKHOLE_BQ_info + || GET_INFO(updatee_bypass) == CAF_BLACKHOLE_info + ) { /* Sigh. It has one. Don't lose those threads! */ if (GET_INFO(updatee_keep) == BLACKHOLE_BQ_info) { /* Urgh. Two queues. Merge them. */ @@ -2673,7 +3422,11 @@ threadSqueezeStack(StgTSO *tso) #endif TICK_UPD_SQUEEZED(); - UPD_IND(updatee_bypass, updatee_keep); /* this wakes the threads up */ + /* 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 + */ + UPD_IND_NOLOCK(updatee_bypass, updatee_keep); /* this wakes the threads up */ sp = (P_)frame - 1; /* sp = stuff to slide */ displacement += sizeofW(StgUpdateFrame); @@ -2686,8 +3439,12 @@ threadSqueezeStack(StgTSO *tso) */ if (is_update_frame) { StgBlockingQueue *bh = (StgBlockingQueue *)frame->updatee; - if (bh->header.info != &BLACKHOLE_BQ_info && + if (bh->header.info != &BLACKHOLE_info && + bh->header.info != &BLACKHOLE_BQ_info && bh->header.info != &CAF_BLACKHOLE_info) { +#if (!defined(LAZY_BLACKHOLING)) && defined(DEBUG) + fprintf(stderr,"Unexpected lazy BHing required at 0x%04x\n",(int)bh); +#endif SET_INFO(bh,&BLACKHOLE_info); } } @@ -2707,9 +3464,10 @@ threadSqueezeStack(StgTSO *tso) else next_frame_bottom = tso->sp - 1; -#if 0 /* DEBUG */ - fprintf(stderr, "sliding [%p, %p] by %ld\n", sp, next_frame_bottom, - displacement); +#if DEBUG + IF_DEBUG(gc, + fprintf(stderr, "sliding [%p, %p] by %ld\n", sp, next_frame_bottom, + displacement)) #endif while (sp >= next_frame_bottom) { @@ -2723,8 +3481,16 @@ threadSqueezeStack(StgTSO *tso) tso->sp += displacement; tso->su = prev_frame; +#if DEBUG + IF_DEBUG(gc, + fprintf(stderr, "@@ threadSqueezeStack: squeezed %d update-frames; found %d BHs; found %d update-, %d stop-, %d catch, %d seq-frames\n", + squeezes, bhs, upd_frames, stop_frames, catch_frames, seq_frames)) +#endif } +//@node Pausing a thread, Index, Stack squeezing +//@subsection Pausing a thread + /* ----------------------------------------------------------------------------- * Pausing a thread * @@ -2732,7 +3498,7 @@ threadSqueezeStack(StgTSO *tso) * here. We also take the opportunity to do stack squeezing if it's * turned on. * -------------------------------------------------------------------------- */ - +//@cindex threadPaused void threadPaused(StgTSO *tso) { @@ -2741,3 +3507,97 @@ threadPaused(StgTSO *tso) else threadLazyBlackHole(tso); } + +/* ----------------------------------------------------------------------------- + * Debugging + * -------------------------------------------------------------------------- */ + +#if DEBUG +//@cindex printMutOnceList +void +printMutOnceList(generation *gen) +{ + StgMutClosure *p, *next; + + p = gen->mut_once_list; + next = p->mut_link; + + fprintf(stderr, "@@ Mut once list %p: ", gen->mut_once_list); + for (; p != END_MUT_LIST; p = next, next = p->mut_link) { + fprintf(stderr, "%p (%s), ", + p, info_type((StgClosure *)p)); + } + fputc('\n', stderr); +} + +//@cindex printMutableList +void +printMutableList(generation *gen) +{ + StgMutClosure *p, *next; + + p = gen->mut_list; + next = p->mut_link; + + fprintf(stderr, "@@ Mutable list %p: ", gen->mut_list); + for (; p != END_MUT_LIST; p = next, next = p->mut_link) { + fprintf(stderr, "%p (%s), ", + p, info_type((StgClosure *)p)); + } + fputc('\n', stderr); +} + +//@cindex maybeLarge +static inline rtsBool +maybeLarge(StgClosure *closure) +{ + StgInfoTable *info = get_itbl(closure); + + /* closure types that may be found on the new_large_objects list; + see scavenge_large */ + return (info->type == MUT_ARR_PTRS || + info->type == MUT_ARR_PTRS_FROZEN || + info->type == TSO || + info->type == ARR_WORDS || + info->type == BCO); +} + + +#endif /* DEBUG */ + +//@node Index, , Pausing a thread +//@subsection Index + +//@index +//* GarbageCollect:: @cindex\s-+GarbageCollect +//* MarkRoot:: @cindex\s-+MarkRoot +//* RevertCAFs:: @cindex\s-+RevertCAFs +//* addBlock:: @cindex\s-+addBlock +//* cleanup_weak_ptr_list:: @cindex\s-+cleanup_weak_ptr_list +//* copy:: @cindex\s-+copy +//* copyPart:: @cindex\s-+copyPart +//* evacuate:: @cindex\s-+evacuate +//* evacuate_large:: @cindex\s-+evacuate_large +//* gcCAFs:: @cindex\s-+gcCAFs +//* isAlive:: @cindex\s-+isAlive +//* maybeLarge:: @cindex\s-+maybeLarge +//* mkMutCons:: @cindex\s-+mkMutCons +//* printMutOnceList:: @cindex\s-+printMutOnceList +//* printMutableList:: @cindex\s-+printMutableList +//* relocate_TSO:: @cindex\s-+relocate_TSO +//* scavenge:: @cindex\s-+scavenge +//* scavenge_large:: @cindex\s-+scavenge_large +//* scavenge_mut_once_list:: @cindex\s-+scavenge_mut_once_list +//* scavenge_mutable_list:: @cindex\s-+scavenge_mutable_list +//* scavenge_one:: @cindex\s-+scavenge_one +//* scavenge_srt:: @cindex\s-+scavenge_srt +//* scavenge_stack:: @cindex\s-+scavenge_stack +//* scavenge_static:: @cindex\s-+scavenge_static +//* threadLazyBlackHole:: @cindex\s-+threadLazyBlackHole +//* threadPaused:: @cindex\s-+threadPaused +//* threadSqueezeStack:: @cindex\s-+threadSqueezeStack +//* traverse_weak_ptr_list:: @cindex\s-+traverse_weak_ptr_list +//* upd_evacuee:: @cindex\s-+upd_evacuee +//* zero_mutable_list:: @cindex\s-+zero_mutable_list +//* zero_static_object_list:: @cindex\s-+zero_static_object_list +//@end index