X-Git-Url: http://git.megacz.com/?p=ghc-hetmet.git;a=blobdiff_plain;f=rts%2FUpdates.h;h=de9276c90908ca29a559038ca73291cf53bd01a8;hp=f20b0884324043d790673e178003b4f100fca2f2;hb=5d52d9b64c21dcf77849866584744722f8121389;hpb=081e9e53205935280f1800f113760a8a72498a6f diff --git a/rts/Updates.h b/rts/Updates.h index f20b088..de9276c 100644 --- a/rts/Updates.h +++ b/rts/Updates.h @@ -9,162 +9,12 @@ #ifndef UPDATES_H #define UPDATES_H -/* ----------------------------------------------------------------------------- - Updates - - We have two layers of update macros. The top layer, UPD_IND() and - friends perform all the work of an update. In detail: - - - if the closure being updated is a blocking queue, then all the - threads waiting on the blocking queue are updated. - - - then the lower level updateWithIndirection() macro is invoked - to actually replace the closure with an indirection (see below). - - -------------------------------------------------------------------------- */ - -#ifdef TICKY_TICKY -# define UPD_IND(updclosure, heapptr) \ - UPD_PERM_IND(updclosure,heapptr) -# define UPD_SPEC_IND(updclosure, ind_info, heapptr, and_then) \ - UPD_PERM_IND(updclosure,heapptr); and_then -#else -# define SEMI ; -# define UPD_IND(updclosure, heapptr) \ - UPD_REAL_IND(updclosure,INFO_PTR(stg_IND_info),heapptr,SEMI) -# define UPD_SPEC_IND(updclosure, ind_info, heapptr, and_then) \ - UPD_REAL_IND(updclosure,ind_info,heapptr,and_then) -#endif - -/* These macros have to work in both C and C--, so here's the - * impedence matching: - */ -#ifdef CMINUSMINUS -#define BLOCK_BEGIN -#define BLOCK_END -#define DECLARE_IPTR(info) W_ info -#define FCALL foreign "C" -#define INFO_PTR(info) info -#define ARG_PTR "ptr" -#else -#define BLOCK_BEGIN { -#define BLOCK_END } -#define DECLARE_IPTR(info) const StgInfoTable *(info) -#define FCALL /* nothing */ -#define INFO_PTR(info) &info -#define StgBlockingQueue_blocking_queue(closure) \ - (((StgBlockingQueue *)closure)->blocking_queue) -#define ARG_PTR /* nothing */ -#endif - -/* UPD_IND actually does a PERM_IND if TICKY_TICKY is on; - if you *really* need an IND use UPD_REAL_IND - */ -#define UPD_REAL_IND(updclosure, ind_info, heapptr, and_then) \ - BLOCK_BEGIN \ - DECLARE_IPTR(info); \ - info = GET_INFO(updclosure); \ - updateWithIndirection(ind_info, \ - updclosure, \ - heapptr, \ - and_then); \ - BLOCK_END - -#if defined(PROFILING) || defined(TICKY_TICKY) -#define UPD_PERM_IND(updclosure, heapptr) \ - BLOCK_BEGIN \ - updateWithPermIndirection(updclosure, \ - heapptr); \ - BLOCK_END -#endif - -#if defined(RTS_SUPPORTS_THREADS) - -# ifdef TICKY_TICKY -# define UPD_IND_NOLOCK(updclosure, heapptr) \ - BLOCK_BEGIN \ - updateWithPermIndirection(updclosure, \ - heapptr); \ - BLOCK_END -# else -# define UPD_IND_NOLOCK(updclosure, heapptr) \ - BLOCK_BEGIN \ - updateWithIndirection(INFO_PTR(stg_IND_info), \ - updclosure, \ - heapptr,); \ - BLOCK_END -# endif - -#else -#define UPD_IND_NOLOCK(updclosure,heapptr) UPD_IND(updclosure,heapptr) +#ifndef CMINUSMINUS +BEGIN_RTS_PRIVATE #endif /* ----------------------------------------------------------------------------- - Awaken any threads waiting on a blocking queue (BLACKHOLE_BQ). - -------------------------------------------------------------------------- */ - -#if defined(PAR) - -/* - In a parallel setup several types of closures might have a blocking queue: - BLACKHOLE_BQ ... same as in the default concurrent setup; it will be - reawakened via calling UPD_IND on that closure after - having finished the computation of the graph - FETCH_ME_BQ ... a global indirection (FETCH_ME) may be entered by a - local TSO, turning it into a FETCH_ME_BQ; it will be - reawakened via calling processResume - RBH ... a revertible black hole may be entered by another - local TSO, putting it onto its blocking queue; since - RBHs only exist while the corresponding closure is in - transit, they will be reawakened via calling - convertToFetchMe (upon processing an ACK message) - - In a parallel setup a blocking queue may contain 3 types of closures: - TSO ... as in the default concurrent setup - BLOCKED_FETCH ... indicating that a TSO on another PE is waiting for - the result of the current computation - CONSTR ... an RBHSave closure (which contains data ripped out of - the closure to make room for a blocking queue; since - it only contains data we use the exisiting type of - a CONSTR closure); this closure is the end of a - blocking queue for an RBH closure; it only exists in - this kind of blocking queue and must be at the end - of the queue -*/ -extern void awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node); -#define DO_AWAKEN_BQ(bqe, node) STGCALL2(awakenBlockedQueue, bqe, node); - -#define AWAKEN_BQ(info,closure) \ - if (info == &stg_BLACKHOLE_BQ_info || \ - info == &stg_FETCH_ME_BQ_info || \ - get_itbl(closure)->type == RBH) { \ - DO_AWAKEN_BQ(((StgBlockingQueue *)closure)->blocking_queue, closure); \ - } - -#elif defined(GRAN) - -extern void awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node); -#define DO_AWAKEN_BQ(bq, node) STGCALL2(awakenBlockedQueue, bq, node); - -/* In GranSim we don't have FETCH_ME or FETCH_ME_BQ closures, so they are - not checked. The rest of the code is the same as for GUM. -*/ -#define AWAKEN_BQ(info,closure) \ - if (info == &stg_BLACKHOLE_BQ_info || \ - get_itbl(closure)->type == RBH) { \ - DO_AWAKEN_BQ(((StgBlockingQueue *)closure)->blocking_queue, closure); \ - } - -#endif /* GRAN || PAR */ - - -/* ----------------------------------------------------------------------------- - Updates: lower-level macros which update a closure with an - indirection to another closure. - - There are several variants of this code. - - PROFILING: + Updates -------------------------------------------------------------------------- */ /* LDV profiling: @@ -198,8 +48,7 @@ extern void awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node); W_ sz; \ W_ i; \ inf = %GET_STD_INFO(p); \ - if (%INFO_TYPE(inf) != HALF_W_(BLACKHOLE) \ - && %INFO_TYPE(inf) != HALF_W_(CAF_BLACKHOLE)) { \ + if (%INFO_TYPE(inf) != HALF_W_(BLACKHOLE)) { \ if (%INFO_TYPE(inf) == HALF_W_(THUNK_SELECTOR)) { \ sz = BYTES_TO_WDS(SIZEOF_StgSelector_NoThunkHdr); \ } else { \ @@ -232,7 +81,10 @@ FILL_SLOP(StgClosure *p) switch (inf->type) { case BLACKHOLE: - case CAF_BLACKHOLE: + goto no_slop; + // we already filled in the slop when we overwrote the thunk + // with BLACKHOLE, and also an evacuated BLACKHOLE is only the + // size of an IND. case THUNK_SELECTOR: sz = sizeofW(StgSelector) - sizeofW(StgThunkHeader); break; @@ -249,6 +101,8 @@ FILL_SLOP(StgClosure *p) for (i = 0; i < sz; i++) { ((StgThunk *)p)->payload[i] = 0; } +no_slop: + ; } #endif /* CMINUSMINUS */ @@ -270,96 +124,56 @@ FILL_SLOP(StgClosure *p) * invert the optimisation. Grrrr --SDM). */ #ifdef CMINUSMINUS -#define generation(n) (W_[generations] + n*SIZEOF_generation) -#define updateWithIndirection(ind_info, p1, p2, and_then) \ + +#define updateWithIndirection(p1, p2, and_then) \ W_ bd; \ \ DEBUG_FILL_SLOP(p1); \ LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(p1); \ StgInd_indirectee(p1) = p2; \ - foreign "C" wb() []; \ + prim %write_barrier() []; \ + SET_INFO(p1, stg_BLACKHOLE_info); \ + LDV_RECORD_CREATE(p1); \ bd = Bdescr(p1); \ - if (bdescr_gen_no(bd) != 0 :: CInt) { \ - foreign "C" recordMutableCap(p1 "ptr", \ - MyCapability() "ptr", \ - bdescr_gen_no(bd)) [R1]; \ - SET_INFO(p1, stg_IND_OLDGEN_info); \ - LDV_RECORD_CREATE(p1); \ + if (bdescr_gen_no(bd) != 0 :: bits16) { \ + recordMutableCap(p1, TO_W_(bdescr_gen_no(bd)), R1); \ TICK_UPD_OLD_IND(); \ and_then; \ } else { \ - SET_INFO(p1, ind_info); \ - LDV_RECORD_CREATE(p1); \ TICK_UPD_NEW_IND(); \ and_then; \ } -#else -#define updateWithIndirection(ind_info, p1, p2, and_then) \ - { \ - bdescr *bd; \ - \ - /* cas(p1, 0, &stg_WHITEHOLE_info); */ \ - ASSERT( (P_)p1 != (P_)p2 && !closure_IND(p1) ); \ - DEBUG_FILL_SLOP(p1); \ - LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(p1); \ - ((StgInd *)p1)->indirectee = p2; \ - wb(); \ - bd = Bdescr((P_)p1); \ - if (bd->gen_no != 0) { \ - recordMutableGenLock(p1, &generations[bd->gen_no]); \ - SET_INFO(p1, &stg_IND_OLDGEN_info); \ - TICK_UPD_OLD_IND(); \ - and_then; \ - } else { \ - SET_INFO(p1, ind_info); \ - LDV_RECORD_CREATE(p1); \ - TICK_UPD_NEW_IND(); \ - and_then; \ - } \ - } -#endif -/* The permanent indirection version isn't performance critical. We - * therefore use an inline C function instead of the C-- macro. - */ -#ifndef CMINUSMINUS -INLINE_HEADER void -updateWithPermIndirection(StgClosure *p1, - StgClosure *p2) -{ - bdescr *bd; - - ASSERT( p1 != p2 && !closure_IND(p1) ); - - /* - * @LDV profiling - * Destroy the old closure. - * Nb: LDV_* stuff cannot mix with ticky-ticky - */ - LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(p1); +#else /* !CMINUSMINUS */ - bd = Bdescr((P_)p1); - if (bd->gen_no != 0) { - recordMutableGenLock(p1, &generations[bd->gen_no]); - ((StgInd *)p1)->indirectee = p2; - SET_INFO(p1, &stg_IND_OLDGEN_PERM_info); - /* - * @LDV profiling - * We have just created a new closure. - */ - LDV_RECORD_CREATE(p1); - TICK_UPD_OLD_PERM_IND(); - } else { +INLINE_HEADER void updateWithIndirection (Capability *cap, + StgClosure *p1, + StgClosure *p2) +{ + bdescr *bd; + + ASSERT( (P_)p1 != (P_)p2 ); + /* not necessarily true: ASSERT( !closure_IND(p1) ); */ + /* occurs in RaiseAsync.c:raiseAsync() */ + DEBUG_FILL_SLOP(p1); + LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(p1); ((StgInd *)p1)->indirectee = p2; - SET_INFO(p1, &stg_IND_PERM_info); - /* - * @LDV profiling - * We have just created a new closure. - */ + write_barrier(); + SET_INFO(p1, &stg_BLACKHOLE_info); LDV_RECORD_CREATE(p1); - TICK_UPD_NEW_PERM_IND(p1); - } + bd = Bdescr((StgPtr)p1); + if (bd->gen_no != 0) { + recordMutableCap(p1, cap, bd->gen_no); + TICK_UPD_OLD_IND(); + } else { + TICK_UPD_NEW_IND(); + } } + +#endif /* CMINUSMINUS */ + +#ifndef CMINUSMINUS +END_RTS_PRIVATE #endif #endif /* UPDATES_H */