rts/Updates.h

   1 /* -----------------------------------------------------------------------------
   2  *
   3  * (c) The GHC Team, 1998-2004
   4  *
   5  * Performing updates.
   6  *
   7  * ---------------------------------------------------------------------------*/
   8
   9 #ifndef UPDATES_H
  10 #define UPDATES_H
  11
  12 /* -----------------------------------------------------------------------------
  13    Updates
  14
  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:
  17
  18       - if the closure being updated is a blocking queue, then all the
  19         threads waiting on the blocking queue are updated.
  20
  21       - then the lower level updateWithIndirection() macro is invoked
  22         to actually replace the closure with an indirection (see below).
  23
  24    -------------------------------------------------------------------------- */
  25
  26 #ifdef TICKY_TICKY
  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
  31 #else
  32 #  define SEMI ;
  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)
  37 #endif
  38
  39 /* These macros have to work in both C and C--, so here's the
  40  * impedence matching:
  41  */
  42 #ifdef CMINUSMINUS
  43 #define BLOCK_BEGIN
  44 #define BLOCK_END
  45 #define DECLARE_IPTR(info)  W_ info
  46 #define FCALL               foreign "C"
  47 #define INFO_PTR(info)      info
  48 #define ARG_PTR             "ptr"
  49 #else
  50 #define BLOCK_BEGIN         {
  51 #define BLOCK_END           }
  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 */
  58 #endif
  59
  60 /* UPD_IND actually does a PERM_IND if TICKY_TICKY is on;
  61    if you *really* need an IND use UPD_REAL_IND
  62  */
  63 #define UPD_REAL_IND(updclosure, ind_info, heapptr, and_then)   \
  64         BLOCK_BEGIN                                             \
  65         DECLARE_IPTR(info);                                     \
  66         info = GET_INFO(updclosure);                            \
  67         updateWithIndirection(ind_info,                         \
  68                               updclosure,                       \
  69                               heapptr,                          \
  70                               and_then);                        \
  71         BLOCK_END
  72
  73 #if defined(PROFILING) || defined(TICKY_TICKY)
  74 #define UPD_PERM_IND(updclosure, heapptr)       \
  75         BLOCK_BEGIN                             \
  76         updateWithPermIndirection(updclosure,   \
  77                                   heapptr);     \
  78         BLOCK_END
  79 #endif
  80
  81 #if defined(RTS_SUPPORTS_THREADS)
  82
  83 # ifdef TICKY_TICKY
  84 #  define UPD_IND_NOLOCK(updclosure, heapptr)   \
  85         BLOCK_BEGIN                             \
  86         updateWithPermIndirection(updclosure,   \
  87                                   heapptr);     \
  88         BLOCK_END
  89 # else
  90 #  define UPD_IND_NOLOCK(updclosure, heapptr)                   \
  91         BLOCK_BEGIN                                             \
  92         updateWithIndirection(INFO_PTR(stg_IND_info),           \
  93                               updclosure,                       \
  94                               heapptr,);                        \
  95         BLOCK_END
  96 # endif
  97
  98 #else
  99 #define UPD_IND_NOLOCK(updclosure,heapptr) UPD_IND(updclosure,heapptr)
 100 #endif
 101
 102 /* -----------------------------------------------------------------------------
 103    Awaken any threads waiting on a blocking queue (BLACKHOLE_BQ).
 104    -------------------------------------------------------------------------- */
 105
 106 #if defined(PAR)
 107
 108 /*
 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)
 121
 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
 132                        of the queue
 133 */
 134 extern void awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node);
 135 #define DO_AWAKEN_BQ(bqe, node)  STGCALL2(awakenBlockedQueue, bqe, node);
 136
 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);                           \
 142         }
 143
 144 #elif defined(GRAN)
 145
 146 extern void awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node);
 147 #define DO_AWAKEN_BQ(bq, node)  STGCALL2(awakenBlockedQueue, bq, node);
 148
 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.
 151 */
 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);                           \
 156         }
 157
 158 #endif /* GRAN || PAR */
 159
 160
 161 /* -----------------------------------------------------------------------------
 162    Updates: lower-level macros which update a closure with an
 163    indirection to another closure.
 164
 165    There are several variants of this code.
 166
 167        PROFILING:
 168    -------------------------------------------------------------------------- */
 169
 170 /* LDV profiling:
 171  * We call LDV_recordDead_FILL_SLOP_DYNAMIC(p1) regardless of the generation in
 172  * which p1 resides.
 173  *
 174  * Note:
 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().
 179  */
 180
 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.
 183  *
 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
 187  * otherwise).
 188  *
 189  * NB. We do *not* do this in THREADED_RTS 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!).
 193  */
 194 #ifdef CMINUSMINUS
 195
 196 #define FILL_SLOP(p)                                                    \
 197   W_ inf;                                                               \
 198   W_ sz;                                                                \
 199   W_ i;                                                                 \
 200   inf = %GET_STD_INFO(p);                                               \
 201   if (%INFO_TYPE(inf) != HALF_W_(THUNK_SELECTOR)                        \
 202         && %INFO_TYPE(inf) != HALF_W_(BLACKHOLE)                        \
 203         && %INFO_TYPE(inf) != HALF_W_(CAF_BLACKHOLE)) {                 \
 204       if (%INFO_TYPE(inf) == HALF_W_(AP_STACK)) {                       \
 205           sz = StgAP_STACK_size(p) + BYTES_TO_WDS(SIZEOF_StgAP_STACK_NoThunkHdr); \
 206       } else {                                                          \
 207           if (%INFO_TYPE(inf) == HALF_W_(AP)) {                         \
 208               sz = TO_W_(StgAP_n_args(p)) +  BYTES_TO_WDS(SIZEOF_StgAP_NoThunkHdr);     \
 209           } else {                                                      \
 210               sz = TO_W_(%INFO_PTRS(inf)) + TO_W_(%INFO_NPTRS(inf));    \
 211           }                                                             \
 212       }                                                                 \
 213       i = 0;                                                            \
 214       for:                                                              \
 215         if (i < sz) {                                                   \
 216           StgThunk_payload(p,i) = 0;                                    \
 217           i = i + 1;                                                    \
 218           goto for;                                                     \
 219         }                                                               \
 220   }
 221
 222 #else /* !CMINUSMINUS */
 223
 224 INLINE_HEADER void
 225 FILL_SLOP(StgClosure *p)
 226 {
 227     StgInfoTable *inf = get_itbl(p);
 228     nat i, sz;
 229
 230     switch (inf->type) {
 231     case BLACKHOLE:
 232     case CAF_BLACKHOLE:
 233     case THUNK_SELECTOR:
 234         return;
 235     case AP:
 236         sz = ((StgAP *)p)->n_args + sizeofW(StgAP) - sizeofW(StgThunkHeader);
 237         break;
 238     case AP_STACK:
 239         sz = ((StgAP_STACK *)p)->size + sizeofW(StgAP_STACK) - sizeofW(StgThunkHeader);
 240         break;
 241     default:
 242         sz = inf->layout.payload.ptrs + inf->layout.payload.nptrs;
 243         break;
 244     }
 245     for (i = 0; i < sz; i++) {
 246         ((StgThunk *)p)->payload[i] = 0;
 247     }
 248 }
 249
 250 #endif /* CMINUSMINUS */
 251
 252 #if !defined(DEBUG) || defined(THREADED_RTS)
 253 #define DEBUG_FILL_SLOP(p) /* do nothing */
 254 #else
 255 #define DEBUG_FILL_SLOP(p) FILL_SLOP(p)
 256 #endif
 257
 258 /* We have two versions of this macro (sadly), one for use in C-- code,
 259  * and the other for C.
 260  *
 261  * The and_then argument is a performance hack so that we can paste in
 262  * the continuation code directly.  It helps shave a couple of
 263  * instructions off the common case in the update code, which is
 264  * worthwhile (the update code is often part of the inner loop).
 265  * (except that gcc now appears to common up this code again and
 266  * invert the optimisation.  Grrrr --SDM).
 267  */
 268 #ifdef CMINUSMINUS
 269 #define generation(n) (W_[generations] + n*SIZEOF_generation)
 270 #define updateWithIndirection(ind_info, p1, p2, and_then)       \
 271     W_ bd;                                                      \
 272                                                                 \
 273     DEBUG_FILL_SLOP(p1);                                        \
 274     LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(p1);                      \
 275     StgInd_indirectee(p1) = p2;                                 \
 276     foreign "C" wb() [];                                        \
 277     bd = Bdescr(p1);                                            \
 278     if (bdescr_gen_no(bd) != 0 :: CInt) {                       \
 279       foreign "C" recordMutableCap(p1 "ptr",                    \
 280                                    MyCapability() "ptr",        \
 281                                    bdescr_gen_no(bd)) [R1];     \
 282       SET_INFO(p1, stg_IND_OLDGEN_info);                        \
 283       LDV_RECORD_CREATE(p1);                                    \
 284       TICK_UPD_OLD_IND();                                       \
 285       and_then;                                                 \
 286     } else {                                                    \
 287       SET_INFO(p1, ind_info);                                   \
 288       LDV_RECORD_CREATE(p1);                                    \
 289       TICK_UPD_NEW_IND();                                       \
 290       and_then;                                                 \
 291   }
 292 #else
 293 #define updateWithIndirection(ind_info, p1, p2, and_then)               \
 294   {                                                                     \
 295     bdescr *bd;                                                         \
 296                                                                         \
 297     /* cas(p1, 0, &stg_WHITEHOLE_info); */                              \
 298     ASSERT( (P_)p1 != (P_)p2 && !closure_IND(p1) );                     \
 299     DEBUG_FILL_SLOP(p1);                                                \
 300     LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(p1);                              \
 301     ((StgInd *)p1)->indirectee = p2;                                    \
 302     wb();                                                               \
 303     bd = Bdescr((P_)p1);                                                \
 304     if (bd->gen_no != 0) {                                              \
 305       recordMutableGenLock(p1, &generations[bd->gen_no]);               \
 306       SET_INFO(p1, &stg_IND_OLDGEN_info);                               \
 307       TICK_UPD_OLD_IND();                                               \
 308       and_then;                                                         \
 309     } else {                                                            \
 310       SET_INFO(p1, ind_info);                                           \
 311       LDV_RECORD_CREATE(p1);                                            \
 312       TICK_UPD_NEW_IND();                                               \
 313       and_then;                                                         \
 314     }                                                                   \
 315   }
 316 #endif
 317
 318 /* The permanent indirection version isn't performance critical.  We
 319  * therefore use an inline C function instead of the C-- macro.
 320  */
 321 #ifndef CMINUSMINUS
 322 INLINE_HEADER void
 323 updateWithPermIndirection(StgClosure *p1,
 324                           StgClosure *p2)
 325 {
 326   bdescr *bd;
 327
 328   ASSERT( p1 != p2 && !closure_IND(p1) );
 329
 330   /*
 331    * @LDV profiling
 332    * Destroy the old closure.
 333    * Nb: LDV_* stuff cannot mix with ticky-ticky
 334    */
 335   LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(p1);
 336
 337   bd = Bdescr((P_)p1);
 338   if (bd->gen_no != 0) {
 339     recordMutableGenLock(p1, &generations[bd->gen_no]);
 340     ((StgInd *)p1)->indirectee = p2;
 341     SET_INFO(p1, &stg_IND_OLDGEN_PERM_info);
 342     /*
 343      * @LDV profiling
 344      * We have just created a new closure.
 345      */
 346     LDV_RECORD_CREATE(p1);
 347     TICK_UPD_OLD_PERM_IND();
 348   } else {
 349     ((StgInd *)p1)->indirectee = p2;
 350     SET_INFO(p1, &stg_IND_PERM_info);
 351     /*
 352      * @LDV profiling
 353      * We have just created a new closure.
 354      */
 355     LDV_RECORD_CREATE(p1);
 356     TICK_UPD_NEW_PERM_IND(p1);
 357   }
 358 }
 359 #endif
 360
 361 #endif /* UPDATES_H */