ghc/includes/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    Updates: lower-level macros which update a closure with an
 162    indirection to another closure.
 163
 164    There are several variants of this code.
 165
 166        PROFILING:
 167    -------------------------------------------------------------------------- */
 168
 169 /* LDV profiling:
 170  * We call LDV_recordDead_FILL_SLOP_DYNAMIC(p1) regardless of the generation in
 171  * which p1 resides.
 172  *
 173  * Note:
 174  *   After all, we do *NOT* need to call LDV_RECORD_CREATE() for both IND and
 175  *   IND_OLDGEN closures because they are inherently used. But, it corrupts
 176  *   the invariants that every closure keeps its creation time in the profiling
 177  *  field. So, we call LDV_RECORD_CREATE().
 178  */
 179
 180 /* In the DEBUG case, we also zero out the slop of the old closure,
 181  * so that the sanity checker can tell where the next closure is.
 182  *
 183  * Two important invariants: we should never try to update a closure
 184  * to point to itself, and the closure being updated should not
 185  * already have been updated (the mutable list will get messed up
 186  * otherwise).
 187  */
 188 #if !defined(DEBUG)
 189
 190 #define DEBUG_FILL_SLOP(p) /* nothing */
 191
 192 #else  /* DEBUG */
 193
 194 #ifdef CMINUSMINUS
 195
 196 #define DEBUG_FILL_SLOP(p)                      \
 197   W_ inf;                                       \
 198   W_ np;                                        \
 199   W_ nw;                                        \
 200   W_ i;                                         \
 201   inf = %GET_STD_INFO(p);                       \
 202   np = TO_W_(%INFO_PTRS(inf));                  \
 203   nw = TO_W_(%INFO_NPTRS(inf));                 \
 204   if (%INFO_TYPE(inf) != HALF_W_(THUNK_SELECTOR)) {     \
 205     i = 0;                                      \
 206     for:                                        \
 207       if (i < np + nw) {                        \
 208         StgClosure_payload(p,i) = 0;            \
 209         i = i + 1;                              \
 210         goto for;                               \
 211       }                                         \
 212   }
 213
 214
 215 #else /* !CMINUSMINUS */
 216
 217 INLINE_HEADER void
 218 DEBUG_FILL_SLOP(StgClosure *p)
 219 {
 220     StgInfoTable *inf = get_itbl(p);
 221     nat np = inf->layout.payload.ptrs,
 222         nw = inf->layout.payload.nptrs, i;
 223     if (inf->type != THUNK_SELECTOR) {
 224         for (i = 0; i < np + nw; i++) {
 225             ((StgClosure *)p)->payload[i] = 0;
 226         }
 227     }
 228 }
 229
 230 #endif /* CMINUSMINUS */
 231 #endif /* DEBUG */
 232
 233 /* We have two versions of this macro (sadly), one for use in C-- code,
 234  * and the other for C.
 235  *
 236  * The and_then argument is a performance hack so that we can paste in
 237  * the continuation code directly.  It helps shave a couple of
 238  * instructions off the common case in the update code, which is
 239  * worthwhile (the update code is often part of the inner loop).
 240  * (except that gcc now appears to common up this code again and
 241  * invert the optimisation.  Grrrr --SDM).
 242  */
 243 #ifdef CMINUSMINUS
 244 #define generation(n) (W_[generations] + n*SIZEOF_generation)
 245 #define updateWithIndirection(ind_info, p1, p2, and_then)       \
 246     W_ bd;                                                      \
 247                                                                 \
 248 /*    ASSERT( p1 != p2 && !closure_IND(p1) );                   \
 249  */ LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(p1);                      \
 250     bd = Bdescr(p1);                                            \
 251     if (bdescr_gen_no(bd) == 0 :: CInt) {                       \
 252       StgInd_indirectee(p1) = p2;                               \
 253       SET_INFO(p1, ind_info);                                   \
 254       LDV_RECORD_CREATE(p1);                                    \
 255       TICK_UPD_NEW_IND();                                       \
 256       and_then;                                                 \
 257     } else {                                                    \
 258       DEBUG_FILL_SLOP(p1);                                      \
 259       foreign "C" recordMutableGen(p1 "ptr",                    \
 260                  generation(TO_W_(bdescr_gen_no(bd))) "ptr");   \
 261       StgInd_indirectee(p1) = p2;                               \
 262       SET_INFO(p1, stg_IND_OLDGEN_info);                        \
 263       LDV_RECORD_CREATE(p1);                                    \
 264       TICK_UPD_OLD_IND();                                       \
 265       and_then;                                                 \
 266   }
 267 #else
 268 #define updateWithIndirection(ind_info, p1, p2, and_then)               \
 269   {                                                                     \
 270     bdescr *bd;                                                         \
 271                                                                         \
 272     ASSERT( (P_)p1 != (P_)p2 && !closure_IND(p1) );                     \
 273     LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(p1);                              \
 274     bd = Bdescr((P_)p1);                                                \
 275     if (bd->gen_no == 0) {                                              \
 276       ((StgInd *)p1)->indirectee = p2;                                  \
 277       SET_INFO(p1, ind_info);                                           \
 278       LDV_RECORD_CREATE(p1);                                            \
 279       TICK_UPD_NEW_IND();                                               \
 280       and_then;                                                         \
 281     } else {                                                            \
 282       DEBUG_FILL_SLOP(p1);                                              \
 283       recordMutableGen(p1, &generations[bd->gen_no]);                   \
 284       ((StgInd *)p1)->indirectee = p2;                                  \
 285       SET_INFO(p1, &stg_IND_OLDGEN_info);                               \
 286       TICK_UPD_OLD_IND();                                               \
 287       and_then;                                                         \
 288     }                                                                   \
 289   }
 290 #endif
 291
 292 /* The permanent indirection version isn't performance critical.  We
 293  * therefore use an inline C function instead of the C-- macro.
 294  */
 295 #ifndef CMINUSMINUS
 296 INLINE_HEADER void
 297 updateWithPermIndirection(StgClosure *p1,
 298                           StgClosure *p2)
 299 {
 300   bdescr *bd;
 301
 302   ASSERT( p1 != p2 && !closure_IND(p1) );
 303
 304   /*
 305    * @LDV profiling
 306    * Destroy the old closure.
 307    * Nb: LDV_* stuff cannot mix with ticky-ticky
 308    */
 309   LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC(p1);
 310
 311   bd = Bdescr((P_)p1);
 312   if (bd->gen_no == 0) {
 313     ((StgInd *)p1)->indirectee = p2;
 314     SET_INFO(p1, &stg_IND_PERM_info);
 315     /*
 316      * @LDV profiling
 317      * We have just created a new closure.
 318      */
 319     LDV_RECORD_CREATE(p1);
 320     TICK_UPD_NEW_PERM_IND(p1);
 321   } else {
 322     recordMutableGen(p1, &generations[bd->gen_no]);
 323     ((StgInd *)p1)->indirectee = p2;
 324     SET_INFO(p1, &stg_IND_OLDGEN_PERM_info);
 325     /*
 326      * @LDV profiling
 327      * We have just created a new closure.
 328      */
 329     LDV_RECORD_CREATE(p1);
 330     TICK_UPD_OLD_PERM_IND();
 331   }
 332 }
 333 #endif
 334
 335 #endif /* UPDATES_H */