1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team, 1998-2004
7 * ---------------------------------------------------------------------------*/
12 /* -----------------------------------------------------------------------------
15 We have two layers of update macros. The top layer, UPD_IND() and
16 friends perform all the work of an update. In detail:
18 - if the closure being updated is a blocking queue, then all the
19 threads waiting on the blocking queue are updated.
21 - then the lower level updateWithIndirection() macro is invoked
22 to actually replace the closure with an indirection (see below).
24 -------------------------------------------------------------------------- */
27 # define UPD_IND(updclosure, heapptr) \
28 UPD_PERM_IND(updclosure,heapptr)
29 # define UPD_SPEC_IND(updclosure, ind_info, heapptr, and_then) \
30 UPD_PERM_IND(updclosure,heapptr); and_then
33 # define UPD_IND(updclosure, heapptr) \
34 UPD_REAL_IND(updclosure,INFO_PTR(stg_IND_info),heapptr,SEMI)
35 # define UPD_SPEC_IND(updclosure, ind_info, heapptr, and_then) \
36 UPD_REAL_IND(updclosure,ind_info,heapptr,and_then)
39 /* These macros have to work in both C and C--, so here's the
45 #define DECLARE_IPTR(info) W_ info
46 #define FCALL foreign "C"
47 #define INFO_PTR(info) info
52 #define DECLARE_IPTR(info) const StgInfoTable *(info)
53 #define FCALL /* nothing */
54 #define INFO_PTR(info) &info
55 #define StgBlockingQueue_blocking_queue(closure) \
56 (((StgBlockingQueue *)closure)->blocking_queue)
57 #define ARG_PTR /* nothing */
60 /* UPD_IND actually does a PERM_IND if TICKY_TICKY is on;
61 if you *really* need an IND use UPD_REAL_IND
63 #define UPD_REAL_IND(updclosure, ind_info, heapptr, and_then) \
66 info = GET_INFO(updclosure); \
67 updateWithIndirection(ind_info, \
73 #if defined(PROFILING) || defined(TICKY_TICKY)
74 #define UPD_PERM_IND(updclosure, heapptr) \
76 updateWithPermIndirection(updclosure, \
81 #if defined(RTS_SUPPORTS_THREADS)
84 # define UPD_IND_NOLOCK(updclosure, heapptr) \
86 updateWithPermIndirection(updclosure, \
90 # define UPD_IND_NOLOCK(updclosure, heapptr) \
92 updateWithIndirection(INFO_PTR(stg_IND_info), \
99 #define UPD_IND_NOLOCK(updclosure,heapptr) UPD_IND(updclosure,heapptr)
102 /* -----------------------------------------------------------------------------
103 Awaken any threads waiting on a blocking queue (BLACKHOLE_BQ).
104 -------------------------------------------------------------------------- */
109 In a parallel setup several types of closures might have a blocking queue:
110 BLACKHOLE_BQ ... same as in the default concurrent setup; it will be
111 reawakened via calling UPD_IND on that closure after
112 having finished the computation of the graph
113 FETCH_ME_BQ ... a global indirection (FETCH_ME) may be entered by a
114 local TSO, turning it into a FETCH_ME_BQ; it will be
115 reawakened via calling processResume
116 RBH ... a revertible black hole may be entered by another
117 local TSO, putting it onto its blocking queue; since
118 RBHs only exist while the corresponding closure is in
119 transit, they will be reawakened via calling
120 convertToFetchMe (upon processing an ACK message)
122 In a parallel setup a blocking queue may contain 3 types of closures:
123 TSO ... as in the default concurrent setup
124 BLOCKED_FETCH ... indicating that a TSO on another PE is waiting for
125 the result of the current computation
126 CONSTR ... an RBHSave closure (which contains data ripped out of
127 the closure to make room for a blocking queue; since
128 it only contains data we use the exisiting type of
129 a CONSTR closure); this closure is the end of a
130 blocking queue for an RBH closure; it only exists in
131 this kind of blocking queue and must be at the end
134 extern void awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node);
135 #define DO_AWAKEN_BQ(bqe, node) STGCALL2(awakenBlockedQueue, bqe, node);
137 #define AWAKEN_BQ(info,closure) \
138 if (info == &stg_BLACKHOLE_BQ_info || \
139 info == &stg_FETCH_ME_BQ_info || \
140 get_itbl(closure)->type == RBH) { \
141 DO_AWAKEN_BQ(((StgBlockingQueue *)closure)->blocking_queue, closure); \
146 extern void awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node);
147 #define DO_AWAKEN_BQ(bq, node) STGCALL2(awakenBlockedQueue, bq, node);
149 /* In GranSim we don't have FETCH_ME or FETCH_ME_BQ closures, so they are
150 not checked. The rest of the code is the same as for GUM.
152 #define AWAKEN_BQ(info,closure) \
153 if (info == &stg_BLACKHOLE_BQ_info || \
154 get_itbl(closure)->type == RBH) { \
155 DO_AWAKEN_BQ(((StgBlockingQueue *)closure)->blocking_queue, closure); \
158 #endif /* GRAN || PAR */
161 /* -----------------------------------------------------------------------------
162 Updates: lower-level macros which update a closure with an
163 indirection to another closure.
165 There are several variants of this code.
168 -------------------------------------------------------------------------- */
171 * We call LDV_recordDead_FILL_SLOP_DYNAMIC(p1) regardless of the generation in
175 * After all, we do *NOT* need to call LDV_RECORD_CREATE() for both IND and
176 * IND_OLDGEN closures because they are inherently used. But, it corrupts
177 * the invariants that every closure keeps its creation time in the profiling
178 * field. So, we call LDV_RECORD_CREATE().
181 /* In the DEBUG case, we also zero out the slop of the old closure,
182 * so that the sanity checker can tell where the next closure is.
184 * Two important invariants: we should never try to update a closure
185 * to point to itself, and the closure being updated should not
186 * already have been updated (the mutable list will get messed up
189 * NB. We do *not* do this in SMP mode, because when we have the
190 * possibility of multiple threads entering the same closure, zeroing
191 * the slop in one of the threads would have a disastrous effect on
192 * the other (seen in the wild!).
194 #if !defined(DEBUG) || defined(SMP)
196 #define DEBUG_FILL_SLOP(p) /* nothing */
202 #define DEBUG_FILL_SLOP(p) \
206 inf = %GET_STD_INFO(p); \
207 if (%INFO_TYPE(inf) != HALF_W_(THUNK_SELECTOR)) { \
208 if (%INFO_TYPE(inf) != HALF_W_(BLACKHOLE)) { \
209 if (%INFO_TYPE(inf) == HALF_W_(AP_STACK)) { \
210 sz = StgAP_STACK_size(p) + BYTES_TO_WDS(SIZEOF_StgAP_STACK_NoHdr); \
212 sz = TO_W_(%INFO_PTRS(inf)) + TO_W_(%INFO_NPTRS(inf)); \
214 i = 1; /* skip over indirectee */ \
217 StgThunk_payload(p,i) = 0; \
223 #else /* !CMINUSMINUS */
226 DEBUG_FILL_SLOP(StgClosure *p)
228 StgInfoTable *inf = get_itbl(p);
236 sz = ((StgAP_STACK *)p)->size + sizeofW(StgAP_STACK) - sizeofW(StgHeader);
239 sz = inf->layout.payload.ptrs + inf->layout.payload.nptrs;
242 // start at one to skip over the indirectee
243 for (i = 1; i < sz; i++) {
244 ((StgThunk *)p)->payload[i] = 0;
248 #endif /* CMINUSMINUS */
251 /* We have two versions of this macro (sadly), one for use in C-- code,
252 * and the other for C.
254 * The and_then argument is a performance hack so that we can paste in
255 * the continuation code directly. It helps shave a couple of
256 * instructions off the common case in the update code, which is
257 * worthwhile (the update code is often part of the inner loop).
258 * (except that gcc now appears to common up this code again and
259 * invert the optimisation. Grrrr --SDM).
262 #define generation(n) (W_[generations] + n*SIZEOF_generation)
263 #define updateWithIndirection(ind_info, p1, p2, and_then) \
266 DEBUG_FILL_SLOP(p1); \
267 LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(p1); \
268 StgInd_indirectee(p1) = p2; \
269 foreign "C" wb() []; \
271 if (bdescr_gen_no(bd) != 0 :: CInt) { \
272 foreign "C" recordMutableCap(p1 "ptr", \
273 MyCapability() "ptr", \
274 bdescr_gen_no(bd)); \
275 SET_INFO(p1, stg_IND_OLDGEN_info); \
276 LDV_RECORD_CREATE(p1); \
277 TICK_UPD_OLD_IND(); \
280 SET_INFO(p1, ind_info); \
281 LDV_RECORD_CREATE(p1); \
282 TICK_UPD_NEW_IND(); \
286 #define updateWithIndirection(ind_info, p1, p2, and_then) \
290 /* cas(p1, 0, &stg_WHITEHOLE_info); */ \
291 ASSERT( (P_)p1 != (P_)p2 && !closure_IND(p1) ); \
292 DEBUG_FILL_SLOP(p1); \
293 LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(p1); \
294 ((StgInd *)p1)->indirectee = p2; \
296 bd = Bdescr((P_)p1); \
297 if (bd->gen_no != 0) { \
298 recordMutableGenLock(p1, &generations[bd->gen_no]); \
299 SET_INFO(p1, &stg_IND_OLDGEN_info); \
300 TICK_UPD_OLD_IND(); \
303 SET_INFO(p1, ind_info); \
304 LDV_RECORD_CREATE(p1); \
305 TICK_UPD_NEW_IND(); \
311 /* The permanent indirection version isn't performance critical. We
312 * therefore use an inline C function instead of the C-- macro.
316 updateWithPermIndirection(StgClosure *p1,
321 ASSERT( p1 != p2 && !closure_IND(p1) );
325 * Destroy the old closure.
326 * Nb: LDV_* stuff cannot mix with ticky-ticky
328 LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(p1);
331 if (bd->gen_no != 0) {
332 recordMutableGenLock(p1, &generations[bd->gen_no]);
333 ((StgInd *)p1)->indirectee = p2;
334 SET_INFO(p1, &stg_IND_OLDGEN_PERM_info);
337 * We have just created a new closure.
339 LDV_RECORD_CREATE(p1);
340 TICK_UPD_OLD_PERM_IND();
342 ((StgInd *)p1)->indirectee = p2;
343 SET_INFO(p1, &stg_IND_PERM_info);
346 * We have just created a new closure.
348 LDV_RECORD_CREATE(p1);
349 TICK_UPD_NEW_PERM_IND(p1);
354 #endif /* UPDATES_H */