Add allocateInGen() for allocating in a specific generation, and cleanups
[ghc.git] / rts / LdvProfile.c
1 /* -----------------------------------------------------------------------------
2 *
3 * (c) The GHC Team, 2001
4 * Author: Sungwoo Park
5 *
6 * Lag/Drag/Void profiling.
7 *
8 * ---------------------------------------------------------------------------*/
9
10 #ifdef PROFILING
11
12 #include "Rts.h"
13 #include "LdvProfile.h"
14 #include "RtsFlags.h"
15 #include "Profiling.h"
16 #include "Stats.h"
17 #include "RtsUtils.h"
18 #include "Schedule.h"
19
20 /* --------------------------------------------------------------------------
21 * Fills in the slop when a *dynamic* closure changes its type.
22 * First calls LDV_recordDead() to declare the closure is dead, and then
23 * fills in the slop.
24 *
25 * Invoked when:
26 * 1) blackholing, UPD_BH_UPDATABLE() and UPD_BH_SINGLE_ENTRY (in
27 * includes/StgMacros.h), threadLazyBlackHole() and
28 * threadSqueezeStack() (in GC.c).
29 * 2) updating with indirection closures, updateWithIndirection()
30 * and updateWithPermIndirection() (in Storage.h).
31 *
32 * LDV_recordDead_FILL_SLOP_DYNAMIC() is not called on 'inherently used'
33 * closures such as TSO. It is not called on PAP because PAP is not updatable.
34 * ----------------------------------------------------------------------- */
35 void
36 LDV_recordDead_FILL_SLOP_DYNAMIC( StgClosure *p )
37 {
38 nat size, i;
39
40 #if defined(__GNUC__) && __GNUC__ < 3 && defined(DEBUG)
41 #error Please use gcc 3.0+ to compile this file with DEBUG; gcc < 3.0 miscompiles it
42 #endif
43
44 if (era > 0) {
45 // very like FILL_SLOP(), except that we call LDV_recordDead().
46 size = closure_sizeW(p);
47
48 LDV_recordDead((StgClosure *)(p), size);
49
50 if (size > sizeofW(StgThunkHeader)) {
51 for (i = 0; i < size - sizeofW(StgThunkHeader); i++) {
52 ((StgThunk *)(p))->payload[i] = 0;
53 }
54 }
55 }
56 }
57
58 /* --------------------------------------------------------------------------
59 * This function is called eventually on every object destroyed during
60 * a garbage collection, whether it is a major garbage collection or
61 * not. If c is an 'inherently used' closure, nothing happens. If c
62 * is an ordinary closure, LDV_recordDead() is called on c with its
63 * proper size which excludes the profiling header portion in the
64 * closure. Returns the size of the closure, including the profiling
65 * header portion, so that the caller can find the next closure.
66 * ----------------------------------------------------------------------- */
67 STATIC_INLINE nat
68 processHeapClosureForDead( StgClosure *c )
69 {
70 nat size;
71 StgInfoTable *info;
72
73 info = get_itbl(c);
74
75 if (info->type != EVACUATED) {
76 ASSERT(((LDVW(c) & LDV_CREATE_MASK) >> LDV_SHIFT) <= era &&
77 ((LDVW(c) & LDV_CREATE_MASK) >> LDV_SHIFT) > 0);
78 ASSERT(((LDVW(c) & LDV_STATE_MASK) == LDV_STATE_CREATE) ||
79 (
80 (LDVW(c) & LDV_LAST_MASK) <= era &&
81 (LDVW(c) & LDV_LAST_MASK) > 0
82 ));
83 }
84
85 if (info->type == EVACUATED) {
86 // The size of the evacuated closure is currently stored in
87 // the LDV field. See SET_EVACUAEE_FOR_LDV() in
88 // includes/StgLdvProf.h.
89 return LDVW(c);
90 }
91
92 size = closure_sizeW(c);
93
94 switch (info->type) {
95 /*
96 'inherently used' cases: do nothing.
97 */
98 case TSO:
99 case MVAR_CLEAN:
100 case MVAR_DIRTY:
101 case MUT_ARR_PTRS_CLEAN:
102 case MUT_ARR_PTRS_DIRTY:
103 case MUT_ARR_PTRS_FROZEN:
104 case MUT_ARR_PTRS_FROZEN0:
105 case ARR_WORDS:
106 case WEAK:
107 case MUT_VAR_CLEAN:
108 case MUT_VAR_DIRTY:
109 case BCO:
110 case STABLE_NAME:
111 case TVAR_WATCH_QUEUE:
112 case TVAR:
113 case TREC_HEADER:
114 case TREC_CHUNK:
115 case INVARIANT_CHECK_QUEUE:
116 case ATOMIC_INVARIANT:
117 return size;
118
119 /*
120 ordinary cases: call LDV_recordDead().
121 */
122 case THUNK:
123 case THUNK_1_0:
124 case THUNK_0_1:
125 case THUNK_SELECTOR:
126 case THUNK_2_0:
127 case THUNK_1_1:
128 case THUNK_0_2:
129 case AP:
130 case PAP:
131 case AP_STACK:
132 case CONSTR:
133 case CONSTR_1_0:
134 case CONSTR_0_1:
135 case CONSTR_2_0:
136 case CONSTR_1_1:
137 case CONSTR_0_2:
138 case FUN:
139 case FUN_1_0:
140 case FUN_0_1:
141 case FUN_2_0:
142 case FUN_1_1:
143 case FUN_0_2:
144 case BLACKHOLE:
145 case SE_BLACKHOLE:
146 case CAF_BLACKHOLE:
147 case SE_CAF_BLACKHOLE:
148 case IND_PERM:
149 case IND_OLDGEN_PERM:
150 /*
151 'Ingore' cases
152 */
153 // Why can we ignore IND/IND_OLDGEN closures? We assume that
154 // any census is preceded by a major garbage collection, which
155 // IND/IND_OLDGEN closures cannot survive. Therefore, it is no
156 // use considering IND/IND_OLDGEN closures in the meanwhile
157 // because they will perish before the next census at any
158 // rate.
159 case IND:
160 case IND_OLDGEN:
161 // Found a dead closure: record its size
162 LDV_recordDead(c, size);
163 return size;
164
165 /*
166 Error case
167 */
168 // static objects
169 case IND_STATIC:
170 case CONSTR_STATIC:
171 case FUN_STATIC:
172 case THUNK_STATIC:
173 case CONSTR_NOCAF_STATIC:
174 // stack objects
175 case UPDATE_FRAME:
176 case CATCH_FRAME:
177 case STOP_FRAME:
178 case RET_DYN:
179 case RET_BCO:
180 case RET_SMALL:
181 case RET_BIG:
182 // others
183 case BLOCKED_FETCH:
184 case FETCH_ME:
185 case FETCH_ME_BQ:
186 case RBH:
187 case REMOTE_REF:
188 case INVALID_OBJECT:
189 default:
190 barf("Invalid object in processHeapClosureForDead(): %d", info->type);
191 return 0;
192 }
193 }
194
195 /* --------------------------------------------------------------------------
196 * Calls processHeapClosureForDead() on every *dead* closures in the
197 * heap blocks starting at bd.
198 * ----------------------------------------------------------------------- */
199 static void
200 processHeapForDead( bdescr *bd )
201 {
202 StgPtr p;
203
204 while (bd != NULL) {
205 p = bd->start;
206 while (p < bd->free) {
207 p += processHeapClosureForDead((StgClosure *)p);
208 while (p < bd->free && !*p) // skip slop
209 p++;
210 }
211 ASSERT(p == bd->free);
212 bd = bd->link;
213 }
214 }
215
216 /* --------------------------------------------------------------------------
217 * Calls processHeapClosureForDead() on every *dead* closures in the nursery.
218 * ----------------------------------------------------------------------- */
219 static void
220 processNurseryForDead( void )
221 {
222 StgPtr p, bdLimit;
223 bdescr *bd;
224
225 bd = MainCapability.r.rNursery->blocks;
226 while (bd->start < bd->free) {
227 p = bd->start;
228 bdLimit = bd->start + BLOCK_SIZE_W;
229 while (p < bd->free && p < bdLimit) {
230 p += processHeapClosureForDead((StgClosure *)p);
231 while (p < bd->free && p < bdLimit && !*p) // skip slop
232 p++;
233 }
234 bd = bd->link;
235 if (bd == NULL)
236 break;
237 }
238 }
239
240 /* --------------------------------------------------------------------------
241 * Calls processHeapClosureForDead() on every *dead* closures in the
242 * small object pool.
243 * ----------------------------------------------------------------------- */
244 static void
245 processSmallObjectPoolForDead( void )
246 {
247 bdescr *bd;
248 StgPtr p;
249
250 bd = g0s0->blocks;
251
252 // first block
253 if (bd == NULL)
254 return;
255
256 p = bd->start;
257 while (p < alloc_Hp) {
258 p += processHeapClosureForDead((StgClosure *)p);
259 while (p < alloc_Hp && !*p) // skip slop
260 p++;
261 }
262 ASSERT(p == alloc_Hp);
263
264 bd = bd->link;
265 while (bd != NULL) {
266 p = bd->start;
267 while (p < bd->free) {
268 p += processHeapClosureForDead((StgClosure *)p);
269 while (p < bd->free && !*p) // skip slop
270 p++;
271 }
272 ASSERT(p == bd->free);
273 bd = bd->link;
274 }
275 }
276
277 /* --------------------------------------------------------------------------
278 * Calls processHeapClosureForDead() on every *dead* closures in the closure
279 * chain.
280 * ----------------------------------------------------------------------- */
281 static void
282 processChainForDead( bdescr *bd )
283 {
284 // Any object still in the chain is dead!
285 while (bd != NULL) {
286 processHeapClosureForDead((StgClosure *)bd->start);
287 bd = bd->link;
288 }
289 }
290
291 /* --------------------------------------------------------------------------
292 * Start a census for *dead* closures, and calls
293 * processHeapClosureForDead() on every closure which died in the
294 * current garbage collection. This function is called from a garbage
295 * collector right before tidying up, when all dead closures are still
296 * stored in the heap and easy to identify. Generations 0 through N
297 * have just beed garbage collected.
298 * ----------------------------------------------------------------------- */
299 void
300 LdvCensusForDead( nat N )
301 {
302 nat g, s;
303
304 // ldvTime == 0 means that LDV profiling is currently turned off.
305 if (era == 0)
306 return;
307
308 if (RtsFlags.GcFlags.generations == 1) {
309 //
310 // Todo: support LDV for two-space garbage collection.
311 //
312 barf("Lag/Drag/Void profiling not supported with -G1");
313 } else {
314 for (g = 0; g <= N; g++)
315 for (s = 0; s < generations[g].n_steps; s++) {
316 if (g == 0 && s == 0) {
317 processSmallObjectPoolForDead();
318 processNurseryForDead();
319 processChainForDead(generations[g].steps[s].large_objects);
320 } else{
321 processHeapForDead(generations[g].steps[s].old_blocks);
322 processChainForDead(generations[g].steps[s].large_objects);
323 }
324 }
325 }
326 }
327
328 /* --------------------------------------------------------------------------
329 * Regard any closure in the current heap as dead or moribund and update
330 * LDV statistics accordingly.
331 * Called from shutdownHaskell() in RtsStartup.c.
332 * Also, stops LDV profiling by resetting ldvTime to 0.
333 * ----------------------------------------------------------------------- */
334 void
335 LdvCensusKillAll( void )
336 {
337 LdvCensusForDead(RtsFlags.GcFlags.generations - 1);
338 }
339
340 #endif /* PROFILING */