1 %************************************************************************
3 \section[CostCentre.lh]{Definitions for Cost Centre Profiling}
5 %************************************************************************
7 Multi-slurp protection:
13 For threaded activity profiling, we need a few bits of the CostCentre
14 environment to be defined, despite the fact that we don't have CostCentre
18 #if defined(PROFILING) || defined(CONCURRENT)
20 # define CC_EXTERN(cc_ident) \
21 extern struct cc CAT2(cc_ident,_struct); \
22 extern CostCentre cc_ident
24 extern CostCentre CCC; /* the current cost centre */
26 extern CostCentre Registered_CC;/* registered cost centre list */
28 CC_EXTERN(CC_MAIN); /* initial MAIN cost centre */
29 CC_EXTERN(CC_GC); /* Garbage Collection cost center */
32 CC_EXTERN(CC_MSG); /* Communications cost center */
33 CC_EXTERN(CC_IDLE); /* Idle-time cost centre */
36 # define REGISTERED_END (CostCentre)4 /* end of list */
37 /* That 4 look likes a HACK, Patrick.
39 # define NOT_REGISTERED (CostCentre)0 /* not yet registered */
43 The compiler declares a static block for each @_scc_@ annotation in the
44 source using the @CC_DECLARE@ macro where @label@, @module@ and
45 @group@ are strings and @ident@ the cost centre identifier.
48 # define CC_IS_CAF 'c'
49 # define CC_IS_DICT 'd'
50 # define CC_IS_SUBSUMED 'S'
51 # define CC_IS_BORING 'B'
53 # define STATIC_CC_REF(cc_ident) &CAT2(cc_ident,_struct)
54 # define DYN_CC_REF(cc_ident) cc_ident /* unused */
56 # define CC_DECLARE(cc_ident,name,module,group,subsumed,is_local) \
57 is_local struct cc CAT2(cc_ident,_struct) \
58 = {NOT_REGISTERED, UNHASHED, name, module, group, \
59 subsumed, INIT_CC_STATS}; \
60 is_local CostCentre cc_ident = STATIC_CC_REF(cc_ident)
62 #endif /* defined(PROFILING) || defined(CONCURRENT) */
65 Definitions relating to the profiling field as a whole.
68 #define PROF_FIXED_HDR (CC_HDR_SIZE)
69 #define PROF_HDR_POSN AFTER_PAR_HDR
70 #define AFTER_PROF_HDR (PROF_FIXED_HDR+PROF_HDR_POSN)
71 #define SET_PROF_HDR(closure,cc) SET_CC_HDR(closure,cc)
72 #define SET_STATIC_PROF_HDR(ident) SET_STATIC_CC_HDR(ident)
75 %************************************************************************
77 \subsection[no-cost-centres]{Dummy definitions if no cost centres}
79 %************************************************************************
81 The cost-centre profiling is only on if the driver turns on
84 These are the {\em dummy} definitions in force if we do {\em NOT}
85 turn on @PROFILING@. Get them out of the way....
88 #if !defined(PROFILING)
90 /*** Declaration Definitions ***/
92 # define CAT_DECLARE(base_name, kind, descr, type)
94 # define CC_HDR_SIZE 0 /* No CC in fixed header */
96 # define SET_CC_HDR(closure, cc) /* Dont set CC header */
97 # define SET_STATIC_CC_HDR(cc) /* No static CC header */
99 # define SET_CCC(cc_ident,do_scc_count)
100 # define SET_DICT_CCC(cc_ident,do_scc_count)
101 # define SET_CCC_RTS(cc_ident,do_sub_count,do_count)
103 # define RESTORE_CCC(cc)
105 # define ENTER_CC_T(cc)
106 # define ENTER_CC_TCL(closure)
107 # define ENTER_CC_F(cc)
108 # define ENTER_CC_FCL(closure)
109 # define ENTER_CC_FSUB()
110 # define ENTER_CC_FCAF(cc)
111 # define ENTER_CC_FLOAD(cc)
112 # define ENTER_CC_PAP(cc)
113 # define ENTER_CC_PAP_CL(closure)
115 /*** Timer and Heap Definitions ***/
117 # define OR_INTERVAL_EXPIRED /* No || as it is false */
119 # define CC_ALLOC(cc, size, kind)
120 # define HEAP_PROFILE_CLOSURE(closure,size)
123 # define START_TIME_PROFILER
124 # define RESTART_TIME_PROFILER
125 # define STOP_TIME_PROFILER
128 #endif /* !defined(PROFILING) */
131 %************************************************************************
133 \subsection[declaring-cost-centres]{Declaring Cost Centres}
135 %************************************************************************
137 Now for the cost-centre profiling stuff.
139 %************************************************************************
141 \subsection[cost-centres]{Location of Cost Centres}
143 %************************************************************************
145 We have a current cost centre, a list of registered cost centres, and
146 an additional cost centre field within the fixed header of all
147 closures. This is adjacent to the info pointer.
150 #if defined(PROFILING)
152 CC_EXTERN(CC_SUBSUMED); /* top level fns SUBSUMED cost centre */
153 CC_EXTERN(CC_OVERHEAD); /* costs due only to profiling machinery */
154 CC_EXTERN(CC_DONTZuCARE); /* placeholder only */
156 CC_EXTERN(CC_CAFs); /* prelude cost centre (CAFs only) */
157 CC_EXTERN(CC_DICTs); /* prelude cost centre (DICTs only) */
159 # define IS_CAF_OR_DICT_CC(cc) \
160 ((cc)->is_subsumed & ' ') /* tests for lower case character */
162 # define IS_SUBSUMED_CC(cc) ((cc)->is_subsumed == CC_IS_SUBSUMED)
166 Definitions referring to the Cost Centre sub-field of the fixed header.
169 # define CC_HDR_SIZE 1 /* words of header */
170 /*R SMinterface.lh */
172 # define CC_HDR_POSN PROF_HDR_POSN /* position in header */
174 # define CC_HDR(closure) (((P_)(closure))[CC_HDR_POSN])
176 # define SET_CC_HDR(closure, cc) \
177 CC_HDR(closure) = (W_)(cc) /* Set closures cost centre */
178 /*R SMinterface.lh (CCC) */
181 For static closures ...
183 # define SET_STATIC_CC_HDR(cc_ident) \
184 , (W_) STATIC_CC_REF(cc_ident) /* static initialisation */
185 /*R SMinterface.lh */
188 The @/*R @{\em file}@ */@ comments indicate that the macro is used
189 regardless in {\em file} so we need a null definition if cost centres
192 %************************************************************************
194 \subsection{Setting the Current Cost Centre}
196 %************************************************************************
198 On execution of an @_scc_@ expression a new cost centre is set.
200 If the new cost centre is different from the CCC we set the CCC and
203 If the cost centre is the same as the CCC no action is required. We do
204 not count the entry to avoid large counts arising from simple
205 recursion. (Huh? WDP 94/07)
208 # define SET_CCC_X(cc,do_subcc_count,do_subdict_count,do_scc_count) \
210 if ((do_subcc_count)) { CCC->sub_scc_count++; } /* inc subcc count of CCC */ \
211 if ((do_subdict_count)) { CCC->sub_dictcc_count++; } /* inc sub dict count of CCC */ \
212 CCC = (CostCentre)(cc); /* set CCC to ident cc */ \
213 ASSERT_IS_REGISTERED(CCC,1); \
214 if ((do_scc_count)) { CCC->scc_count++; } /* inc scc count of new CCC*/ \
217 # define SET_CCC(cc_ident,do_scc_count) \
218 SET_CCC_X(STATIC_CC_REF(cc_ident),do_scc_count,0,do_scc_count)
220 # define SET_DICT_CCC(cc_ident,do_scc_count) \
221 SET_CCC_X(STATIC_CC_REF(cc_ident),0,do_scc_count,do_scc_count)
223 # define SET_CCC_RTS(cc_ident,do_sub_count,do_scc_count) \
224 SET_CCC_X(STATIC_CC_REF(cc_ident),do_sub_count,0,do_scc_count)
227 We have this @RESTORE_CCC@ macro, rather than just an assignment,
228 in case we want to do any paranoia-checking here.
230 # define RESTORE_CCC(cc) \
232 CCC = (CostCentre) (cc); \
233 ASSERT_IS_REGISTERED(CCC,1); \
237 On entry to top level CAFs we count the scc ...
239 # define ENTER_CC_CAF_X(cc) \
241 CCC->sub_cafcc_count++; /* inc subcaf count of CCC */ \
242 CCC = (CostCentre)(cc); /* set CCC to ident cc */ \
243 ASSERT_IS_REGISTERED(CCC,1); \
244 CCC->scc_count++; /* inc scc count of CAF cc */ \
247 # define ENTER_CC_CAF(cc_ident) ENTER_CC_CAF_X(STATIC_CC_REF(cc_ident))
248 # define ENTER_CC_CAF_CL(closure) ENTER_CC_CAF_X((CostCentre)CC_HDR(closure))
251 On entering a closure we only count the enter to thunks ...
253 # define ENTER_CC_T(cc) \
255 CCC = (CostCentre)(cc); \
256 ASSERT_IS_REGISTERED(CCC,1); \
257 CCC_DETAIL_COUNT(CCC->thunk_count); \
260 # define ENTER_CC_TCL(closure) \
261 ENTER_CC_T(CC_HDR(closure))
263 /* Here is our special "hybrid" case when we do *not* set the CCC.
264 (a) The closure is a function, not a thunk;
265 (b) The CC is CAF/DICT-ish.
267 # define ENTER_CC_F(centre) \
269 CostCentre cc = (CostCentre) (centre); \
270 ASSERT_IS_REGISTERED(cc,1); \
271 if ( ! IS_CAF_OR_DICT_CC(cc) ) { \
274 CCC_DETAIL_COUNT(cc->caffun_subsumed); \
275 CCC_DETAIL_COUNT(CCC->subsumed_caf_count); \
277 CCC_DETAIL_COUNT(CCC->function_count); \
280 # define ENTER_CC_FCL(closure) \
281 ENTER_CC_F(CC_HDR(closure))
283 # define ENTER_CC_FSUB() \
285 CCC_DETAIL_COUNT(CCC->subsumed_fun_count); \
286 CCC_DETAIL_COUNT(CCC->function_count); \
289 # define ENTER_CC_FCAF(centre) \
291 CostCentre cc = (CostCentre) (centre); \
292 ASSERT_IS_REGISTERED(cc,1); \
293 CCC_DETAIL_COUNT(cc->caffun_subsumed); \
294 CCC_DETAIL_COUNT(CCC->subsumed_caf_count); \
295 CCC_DETAIL_COUNT(CCC->function_count); \
298 # define ENTER_CC_FLOAD(cc) \
300 CCC = (CostCentre)(cc); \
301 ASSERT_IS_REGISTERED(CCC,1); \
302 CCC_DETAIL_COUNT(CCC->function_count); \
305 /* These ENTER_CC_PAP things are only used in the RTS */
307 # define ENTER_CC_PAP(centre) \
309 CostCentre cc = (CostCentre) (centre); \
310 ASSERT_IS_REGISTERED(cc,1); \
311 if ( ! IS_CAF_OR_DICT_CC(cc) ) { \
314 CCC_DETAIL_COUNT(cc->caffun_subsumed); \
315 CCC_DETAIL_COUNT(CCC->subsumed_caf_count); \
317 CCC_DETAIL_COUNT(CCC->pap_count); \
320 # define ENTER_CC_PAP_CL(closure) \
321 ENTER_CC_PAP(CC_HDR(closure))
323 # if defined(PROFILING_DETAIL_COUNTS)
324 # define CCC_DETAIL_COUNT(inc_this) ((inc_this)++)
326 # define CCC_DETAIL_COUNT(inc_this) /*nothing*/
329 #endif /* PROFILING */
332 %************************************************************************
334 \subsection{``Registering'' cost-centres}
336 %************************************************************************
338 Cost centres are registered at startup by calling a registering
339 routine in each module. Each module registers its cost centres and
340 calls the registering routine for all imported modules. The RTS calls
341 the registering routine for the module Main. This registering must be
342 done before initialisation since the evaluation required for
343 initialisation may use the cost centres.
345 As the code for each module uses tail calls we use an auxiliary stack
346 (in the heap) to record imported modules still to be registered. At
347 the bottom of the stack is NULL which indicates that
348 @miniInterpretEnd@ should be resumed.
350 @START_REGISTER@ and @END_REGISTER@ are special macros used to
351 delimit the function. @END_REGISTER@ pops the next registering
352 routine off the stack and jumps to it. @REGISTER_CC@ registers a cost
353 centre. @REGISTER_IMPORT@ pushes a modules registering routine onto
357 #if defined(PROFILING)
359 extern F_ _regMain (STG_NO_ARGS);
360 extern F_ *register_stack;
362 # define PUSH_REGISTER_STACK(reg_function) \
363 *(register_stack++) = (F_)reg_function
365 # define POP_REGISTER_STACK \
368 # define START_REGISTER_CCS(reg_mod_name) \
369 static int _module_registered = 0; \
370 STGFUN(reg_mod_name) { \
372 if (! _module_registered) { \
373 _module_registered = 1
375 # define START_REGISTER_PRELUDE(reg_mod_name) \
376 static int CAT2(reg_mod_name,_done) = 0; \
377 STGFUN(reg_mod_name) { \
379 if (! CAT2(reg_mod_name,_done)) { \
380 CAT2(reg_mod_name,_done) = 1
382 # define REGISTER_IMPORT(reg_mod_name) \
383 do { extern F_ reg_mod_name (STG_NO_ARGS) ; \
384 PUSH_REGISTER_STACK(reg_mod_name) ; \
387 # define END_REGISTER_CCS() \
390 F_ cont = POP_REGISTER_STACK; \
391 if (cont == NULL) { \
392 RESUME_(miniInterpretEnd); \
399 #endif /* PROFILING */
402 We don't want to attribute costs to an unregistered cost-centre:
404 #if !defined(PROFILING) || !defined(DEBUG)
405 # define ASSERT_IS_REGISTERED(cc,chk_not_overhead) /*nothing*/
407 # define ASSERT_IS_REGISTERED(cc,chk_not_overhead) \
408 do { /* beware of cc name-capture */ \
409 CostCentre c_c = (CostCentre) (cc); \
410 if (c_c->registered == NOT_REGISTERED) { \
411 fprintf(stderr,"Entering unregistered CC: %s\n",c_c->label); \
412 /* unsafe c-call, BTW */ \
414 if ( (chk_not_overhead) && c_c == STATIC_CC_REF(CC_OVERHEAD) ) { \
415 fprintf(stderr,"CC should not be OVERHEAD!: %s\n",c_c->label); \
416 /* unsafe c-call, BTW */ \
420 #define REGISTER_CC(cc) \
422 extern CostCentre cc; \
423 if (((CostCentre)(cc))->registered == NOT_REGISTERED) { \
424 ((CostCentre)(cc))->registered = Registered_CC; \
425 Registered_CC = (CostCentre)(cc); \
430 %************************************************************************
432 \subsection[declaring-closure-categories]{Declaring Closure Categories}
434 %************************************************************************
436 Closure category records are attached to the info table of the
437 closure. They are declared with the info table. Hashing will map
438 similar categories to the same hash value allowing statistics to be
439 grouped by closure category.
441 The declaration macros expand to nothing if cost centre profiling is
444 Note from ADR: Very dubious Malloc Ptr addition -- should probably just
445 reuse @CON_K@ (or something) in runtime/main/StgStartup.lhc.
446 Similarily, the SP stuff should probably be the highly uninformative
449 SOF 4/96: Renamed MallocPtr_K to ForeignObj_K
452 #if defined(PROFILING)
462 # define ForeignObj_K 7 /* Malloc Pointer */
463 # define SPT_K 8 /* Stable Pointer Table */
466 # define INTERNAL_KIND 10
468 typedef struct ClCat {
469 hash_t index_val; /* hashed value */
470 I_ selected; /* is this category selected (-1 == not memoised, selected? 0 or 1) */
471 I_ kind; /* closure kind -- as above */
472 char *descr; /* source derived string detailing closure description */
473 char *type; /* source derived string detailing closure type */
476 /* We put pointers to these ClCat things in info tables.
477 We need these ClCat things because they are mutable,
478 whereas info tables are immutable. (WDP 94/11)
480 We really should not make funny names by appending _CAT.
483 # define MK_CAT_IDENT(i) CAT2(i,_CAT)
484 # define REF_CAT_IDENT(i) (&MK_CAT_IDENT(i))
486 # define CAT_DECLARE(base_name, kind, descr, type) \
487 static struct ClCat MK_CAT_IDENT(base_name) = {UNHASHED,-1,kind,descr,type};
489 #endif /* PROFILING */
492 %************************************************************************
494 \subsection[timer-interupts]{Processing of Timer Signals}
496 %************************************************************************
498 Stuff to do with timer signals:
500 #if defined(PROFILING) || defined(PAR)
502 extern I_ time_profiling; /* Flag indicating if timer/serial profiling is required */
504 extern I_ interval_expired; /* Flag set by signal handler */
505 extern I_ current_interval; /* Current interval number -- used as time stamp */
506 extern I_ interval_ticks; /* No of ticks in an interval */
508 extern I_ previous_ticks; /* ticks in previous intervals */
509 extern I_ current_ticks; /* ticks in current interval */
511 extern void set_time_profile_handler(STG_NO_ARGS);
512 extern void start_time_profiler(STG_NO_ARGS);
513 extern void restart_time_profiler(STG_NO_ARGS);
514 extern void stop_time_profiler(STG_NO_ARGS);
516 # define TICK_FREQUENCY 50 /* ticks per second */
517 # define TICK_MILLISECS (1000/TICK_FREQUENCY) /* milli-seconds per tick */
519 # define DEFAULT_INTERVAL TICK_FREQUENCY /* 1 second */
521 /* These are never called directly from threaded code */
522 # define START_TIME_PROFILER ULTRASAFESTGCALL0(void,(void *),start_time_profiler) /*R StgOverflow.lc */
523 # define RESTART_TIME_PROFILER ULTRASAFESTGCALL0(void,(void *),restart_time_profiler) /*R StgOverflow.lc */
524 # define STOP_TIME_PROFILER ULTRASAFESTGCALL0(void,(void *),stop_time_profiler) /*R StgOverflow.lc */
526 # if defined(PROFILING)
527 # define OR_INTERVAL_EXPIRED || (interval_expired) /*R StgMacros.h */
532 %************************************************************************
534 \subsection[indexing]{Indexing of Cost Centres and Categories}
536 %************************************************************************
538 Cost Centres and Closure Categories are hashed to provide indexes
539 against which arbitrary information can be stored. These indexes are
540 memoised in the appropriate cost centre or category record and
541 subsequent hashes avoided by the index routine (it simply returns the
544 There are different features which can be hashed allowing information
545 to be stored for different groupings. Cost centres have the cost
546 centre recorded (using the pointer), module and group. Closure
547 categories have the closure description and the type
548 description. Records with the same feature will be hashed to the same
551 The initialisation routines, @init_index_<feature>@, allocate a hash
552 table in which the cost centre / category records are stored. The
553 lower bound for the table size is taken from @max_<feature>_no@. They
554 return the actual table size used (the next power of 2). Unused
555 locations in the hash table are indicated by a 0 entry. Successive
556 @init_index_<feature>@ calls just return the actual table size.
558 Calls to @index_<feature>@ will insert the cost centre / category
559 record in the <feature> hash table, if not already inserted. The hash
560 index is memoised in the record and returned.
562 CURRENTLY ONLY ONE MEMOISATION SLOT IS AVILABLE IN EACH RECORD SO
563 HASHING CAN ONLY BE DONE ON ONE FEATURE FOR EACH RECORD. This can be
564 easily relaxed at the expense of extra memoisation space or continued
567 The initialisation routines must be called before initialisation of
568 the stacks and heap as they require to allocate storage. It is also
569 expected that the caller may want to allocate additional storage in
570 which to store profiling information based on the return table size
574 # if defined(PROFILING)
576 # define DEFAULT_MAX_CC 4096
577 # define DEFAULT_MAX_MOD 256
578 # define DEFAULT_MAX_GRP 128
579 # define DEFAULT_MAX_DESCR 4096
580 # define DEFAULT_MAX_TYPE 1024
582 extern hash_t max_cc_no; /* Hash on CC ptr */
583 extern CostCentre *index_cc_table;
584 extern hash_t init_index_cc(STG_NO_ARGS);
585 extern hash_t index_cc PROTO((CostCentre cc));
587 extern hash_t max_mod_no; /* Hash on CC module */
588 extern CostCentre *index_mod_table;
589 extern hash_t init_index_mod(STG_NO_ARGS);
590 extern hash_t index_mod PROTO((CostCentre cc));
592 extern hash_t max_grp_no; /* Hash on CC group */
593 extern CostCentre *index_grp_table;
594 extern hash_t init_index_grp(STG_NO_ARGS);
595 extern hash_t index_grp PROTO((CostCentre cc));
597 extern hash_t max_descr_no; /* Hash on closure description */
598 extern ClCategory *index_descr_table;
599 extern hash_t init_index_descr(STG_NO_ARGS);
600 extern hash_t index_descr PROTO((ClCategory clcat));
602 extern hash_t max_type_no; /* Hash on type description */
603 extern ClCategory *index_type_table;
604 extern hash_t init_index_type(STG_NO_ARGS);
605 extern hash_t index_type PROTO((ClCategory clcat));
607 # endif /* PROFILING */
611 %************************************************************************
613 \subsection[metering]{Metering of Statistics}
615 %************************************************************************
617 @scc_count@ is incremented by the @SetCC@ macro in section
618 \ref{manipulating-cost-centres} above. Below we have the time tick and
622 # define CC_TICK(centre) \
623 do { CostCentre cc = (CostCentre) (centre); \
624 ASSERT_IS_REGISTERED(cc,1); \
625 cc->time_ticks += 1; \
628 # if defined(PROFILING)
629 # define CC_ALLOC(centre, size, kind) \
630 do { CostCentre cc = (CostCentre) (centre); \
631 ASSERT_IS_REGISTERED(cc,0/*OK if OVERHEAD*/); \
632 CCC_DETAIL_COUNT(cc->mem_allocs); \
633 cc->mem_alloc += (size) - (PROF_FIXED_HDR + TICKY_FIXED_HDR); \
639 %************************************************************************
641 \subsection[cost-centre-profiling]{Cost Centre Profiling}
643 %************************************************************************
646 I_ init_cc_profiling PROTO((I_ rts_argc, char *rts_argv[], char *prog_argv[]));
647 void report_cc_profiling PROTO((I_ final));
649 void cc_register(STG_NO_ARGS);
650 void cc_sort PROTO((CostCentre *sort, char sort_on));
651 rtsBool cc_to_ignore PROTO((CostCentre));
654 %************************************************************************
656 \subsection[heap-profiling]{Heap Profiling}
658 %************************************************************************
661 # if defined(PROFILING)
663 I_ heap_profile_init PROTO((char *argv[]));
665 extern void heap_profile_finish(STG_NO_ARGS);
667 extern void heap_profile_setup(STG_NO_ARGS); /* called at start of heap profile */
668 extern void heap_profile_done(STG_NO_ARGS); /* called at end of heap profile */
670 extern void (* heap_profile_fn) PROTO((P_ closure,I_ size));
672 extern I_ earlier_ticks; /* no. of earlier ticks grouped */
673 extern hash_t time_intervals; /* no. of time intervals reported -- 18 */
675 # define HEAP_PROFILE_CLOSURE(closure,size) \
677 if (heap_profile_fn) { \
678 STGCALL2(void,(void *, P_, I_),(*heap_profile_fn),closure,size); \
681 # endif /* PROFILING */
684 End multi-slurp protection:
686 #endif /* PROFILING || PAR */
688 #endif /* CostCentre_H */