2 * Copyright (c) 2001 by Hewlett-Packard Company. All rights reserved.
4 * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
5 * OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
7 * Permission is hereby granted to use or copy this program
8 * for any purpose, provided the above notices are retained on all copies.
9 * Permission to modify the code and to distribute modified code is granted,
10 * provided the above notices are retained, and a notice that the code was
11 * modified is included with the above copyright notice.
16 * This implements a full, though not well-tuned, representation of the
17 * backwards points-to graph. This is used to test for non-GC-robust
18 * data structures; the code is not used during normal garbage collection.
20 * One restriction is that we drop all back-edges from nodes with very
21 * high in-degree, and simply add them add them to a list of such
22 * nodes. They are then treated as permanent roots. Id this by itself
23 * doesn't introduce a space leak, then such nodes can't contribute to
24 * a growing space leak.
27 #ifdef MAKE_BACK_GRAPH
29 #define MAX_IN 10 /* Maximum in-degree we handle directly */
31 #include "private/dbg_mlc.h"
34 #if !defined(DBG_HDRS_ALL) || (ALIGNMENT != CPP_WORDSZ/8) || !defined(UNIX_LIKE)
35 # error Configuration doesnt support MAKE_BACK_GRAPH
38 /* We store single back pointers directly in the object's oh_bg_ptr field. */
39 /* If there is more than one ptr to an object, we store q | FLAG_MANY, */
40 /* where q is a pointer to a back_edges object. */
41 /* Every once in a while we use a back_edges object even for a single */
42 /* pointer, since we need the other fields in the back_edges structure to */
43 /* be present in some fraction of the objects. Otherwise we get serious */
44 /* performance issues. */
47 typedef struct back_edges_struct {
48 word n_edges; /* Number of edges, including those in continuation */
51 # define RETAIN 1 /* Directly points to a reachable object; */
52 /* retain for next GC. */
53 unsigned short height_gc_no;
54 /* If height > 0, then the GC_gc_no value when it */
55 /* was computed. If it was computed this cycle, then */
56 /* it is current. If it was computed during the */
57 /* last cycle, then it represents the old height, */
58 /* which is only saved for live objects referenced by */
59 /* dead ones. This may grow due to refs from newly */
62 /* Longest path through unreachable nodes to this node */
63 /* that we found using depth first search. */
65 # define HEIGHT_UNKNOWN ((signed_word)(-2))
66 # define HEIGHT_IN_PROGRESS ((signed_word)(-1))
68 struct back_edges_struct *cont;
69 /* Pointer to continuation structure; we use only the */
70 /* edges field in the continuation. */
71 /* also used as free list link. */
74 /* Allocate a new back edge structure. Should be more sophisticated */
75 /* if this were production code. */
76 #define MAX_BACK_EDGE_STRUCTS 100000
77 static back_edges *back_edge_space = 0;
78 int GC_n_back_edge_structs = 0; /* Serves as pointer to never used */
79 /* back_edges space. */
80 static back_edges *avail_back_edges = 0;
81 /* Pointer to free list of deallocated */
82 /* back_edges structures. */
84 static back_edges * new_back_edges(void)
86 if (0 == back_edge_space) {
87 back_edge_space = (back_edges *)
88 GET_MEM(MAX_BACK_EDGE_STRUCTS*sizeof(back_edges));
89 GC_add_to_our_memory((ptr_t)back_edge_space,
90 MAX_BACK_EDGE_STRUCTS*sizeof(back_edges));
92 if (0 != avail_back_edges) {
93 back_edges * result = avail_back_edges;
94 avail_back_edges = result -> cont;
98 if (GC_n_back_edge_structs >= MAX_BACK_EDGE_STRUCTS - 1) {
99 ABORT("needed too much space for back edges: adjust "
100 "MAX_BACK_EDGE_STRUCTS");
102 return back_edge_space + (GC_n_back_edge_structs++);
105 /* Deallocate p and its associated continuation structures. */
106 static void deallocate_back_edges(back_edges *p)
108 back_edges *last = p;
110 while (0 != last -> cont) last = last -> cont;
111 last -> cont = avail_back_edges;
112 avail_back_edges = p;
115 /* Table of objects that are currently on the depth-first search */
116 /* stack. Only objects with in-degree one are in this table. */
117 /* Other objects are identified using HEIGHT_IN_PROGRESS. */
118 /* FIXME: This data structure NEEDS IMPROVEMENT. */
119 #define INITIAL_IN_PROGRESS 10000
120 static ptr_t * in_progress_space = 0;
121 static size_t in_progress_size = 0;
122 static size_t n_in_progress = 0;
124 static void push_in_progress(ptr_t p)
126 if (n_in_progress >= in_progress_size)
127 if (in_progress_size == 0) {
128 in_progress_size = INITIAL_IN_PROGRESS;
129 in_progress_space = (ptr_t *)GET_MEM(in_progress_size * sizeof(ptr_t));
130 GC_add_to_our_memory((ptr_t)in_progress_space,
131 in_progress_size * sizeof(ptr_t));
133 ptr_t * new_in_progress_space;
134 in_progress_size *= 2;
135 new_in_progress_space = (ptr_t *)
136 GET_MEM(in_progress_size * sizeof(ptr_t));
137 GC_add_to_our_memory((ptr_t)new_in_progress_space,
138 in_progress_size * sizeof(ptr_t));
139 BCOPY(in_progress_space, new_in_progress_space,
140 n_in_progress * sizeof(ptr_t));
141 in_progress_space = new_in_progress_space;
142 /* FIXME: This just drops the old space. */
144 if (in_progress_space == 0)
145 ABORT("MAKE_BACK_GRAPH: Out of in-progress space: "
146 "Huge linear data structure?");
147 in_progress_space[n_in_progress++] = p;
150 static GC_bool is_in_progress(ptr_t p)
153 for (i = 0; i < n_in_progress; ++i) {
154 if (in_progress_space[i] == p) return TRUE;
159 static void pop_in_progress(ptr_t p)
162 GC_ASSERT(in_progress_space[n_in_progress] == p);
165 #define GET_OH_BG_PTR(p) \
166 (ptr_t)REVEAL_POINTER(((oh *)(p)) -> oh_bg_ptr)
167 #define SET_OH_BG_PTR(p,q) (((oh *)(p)) -> oh_bg_ptr) = HIDE_POINTER(q)
169 /* Execute s once for each predecessor q of p in the points-to graph. */
170 /* s should be a bracketed statement. We declare q. */
171 #define FOR_EACH_PRED(q, p, s) \
173 ptr_t q = GET_OH_BG_PTR(p); \
174 if (!((word)q & FLAG_MANY)) { \
175 if (q && !((word)q & 1)) s \
176 /* !((word)q & 1) checks for a misnterpreted freelist link */ \
178 back_edges *orig_be_ = (back_edges *)((word)q & ~FLAG_MANY); \
179 back_edges *be_ = orig_be_; \
180 int total_, local_; \
181 int n_edges_ = be_ -> n_edges; \
182 for (total_ = 0, local_ = 0; total_ < n_edges_; ++local_, ++total_) { \
183 if (local_ == MAX_IN) { \
187 q = be_ -> edges[local_]; s \
192 /* Ensure that p has a back_edges structure associated with it. */
193 static void ensure_struct(ptr_t p)
195 ptr_t old_back_ptr = GET_OH_BG_PTR(p);
197 if (!((word)old_back_ptr & FLAG_MANY)) {
198 back_edges *be = new_back_edges();
200 if (0 == old_back_ptr) {
204 be -> edges[0] = old_back_ptr;
206 be -> height = HEIGHT_UNKNOWN;
207 be -> height_gc_no = GC_gc_no - 1;
208 GC_ASSERT(be >= back_edge_space);
209 SET_OH_BG_PTR(p, (word)be | FLAG_MANY);
213 /* Add the (forward) edge from p to q to the backward graph. Both p */
214 /* q are pointers to the object base, i.e. pointers to an oh. */
215 static void add_edge(ptr_t p, ptr_t q)
217 ptr_t old_back_ptr = GET_OH_BG_PTR(q);
218 back_edges * be, *be_cont;
220 static unsigned random_number = 13;
221 # define GOT_LUCKY_NUMBER (((++random_number) & 0x7f) == 0)
222 /* A not very random number we use to occasionally allocate a */
223 /* back_edges structure even for a single backward edge. This */
224 /* prevents us from repeatedly tracing back through very long */
225 /* chains, since we will have some place to store height and */
226 /* in_progress flags along the way. */
228 GC_ASSERT(p == GC_base(p) && q == GC_base(q));
229 if (!GC_HAS_DEBUG_INFO(q) || !GC_HAS_DEBUG_INFO(p)) {
230 /* This is really a misinterpreted free list link, since we saw */
231 /* a pointer to a free list. Dont overwrite it! */
234 if (0 == old_back_ptr) {
236 if (GOT_LUCKY_NUMBER) ensure_struct(q);
239 /* Check whether it was already in the list of predecessors. */
240 FOR_EACH_PRED(pred, q, { if (p == pred) return; });
242 old_back_ptr = GET_OH_BG_PTR(q);
243 be = (back_edges *)((word)old_back_ptr & ~FLAG_MANY);
244 for (i = be -> n_edges, be_cont = be; i > MAX_IN;
245 be_cont = be_cont -> cont, i -= MAX_IN) {}
247 be_cont -> cont = new_back_edges();
248 be_cont = be_cont -> cont;
251 be_cont -> edges[i] = p;
253 if (be -> n_edges == 100) {
255 if (GC_print_stats) {
256 GC_err_printf("The following object has in-degree >= 100:\n");
257 GC_print_heap_obj(q);
263 typedef void (*per_object_func)(ptr_t p, size_t n_bytes, word gc_descr);
265 static void per_object_helper(struct hblk *h, word fn)
268 size_t sz = hhdr -> hb_sz;
269 word descr = hhdr -> hb_descr;
270 per_object_func f = (per_object_func)fn;
274 f((ptr_t)(h -> hb_body + i), sz, descr);
276 } while (i + sz <= BYTES_TO_WORDS(HBLKSIZE));
279 void GC_apply_to_each_object(per_object_func f)
281 GC_apply_to_all_blocks(per_object_helper, (word)f);
284 static void reset_back_edge(ptr_t p, size_t n_bytes, word gc_descr)
286 /* Skip any free list links, or dropped blocks */
287 if (GC_HAS_DEBUG_INFO(p)) {
288 ptr_t old_back_ptr = GET_OH_BG_PTR(p);
289 if ((word)old_back_ptr & FLAG_MANY) {
290 back_edges *be = (back_edges *)((word)old_back_ptr & ~FLAG_MANY);
291 if (!(be -> flags & RETAIN)) {
292 deallocate_back_edges(be);
297 GC_ASSERT(GC_is_marked(p));
299 /* Back edges may point to objects that will not be retained. */
300 /* Delete them for now, but remember the height. */
301 /* Some will be added back at next GC. */
303 if (0 != be -> cont) {
304 deallocate_back_edges(be -> cont);
308 GC_ASSERT(GC_is_marked(p));
310 /* We only retain things for one GC cycle at a time. */
311 be -> flags &= ~RETAIN;
313 } else /* Simple back pointer */ {
314 /* Clear to avoid dangling pointer. */
320 static void add_back_edges(ptr_t p, size_t n_bytes, word gc_descr)
322 word *currentp = (word *)(p + sizeof(oh));
324 /* For now, fix up non-length descriptors conservatively. */
325 if((gc_descr & GC_DS_TAGS) != GC_DS_LENGTH) {
328 while (currentp < (word *)(p + gc_descr)) {
329 word current = *currentp++;
330 FIXUP_POINTER(current);
331 if (current >= (word)GC_least_plausible_heap_addr &&
332 current <= (word)GC_greatest_plausible_heap_addr) {
333 ptr_t target = GC_base((void *)current);
341 /* Rebuild the representation of the backward reachability graph. */
342 /* Does not examine mark bits. Can be called before GC. */
343 void GC_build_back_graph(void)
345 GC_apply_to_each_object(add_back_edges);
348 /* Return an approximation to the length of the longest simple path */
349 /* through unreachable objects to p. We refer to this as the height */
351 static word backwards_height(ptr_t p)
354 ptr_t back_ptr = GET_OH_BG_PTR(p);
357 if (0 == back_ptr) return 1;
358 if (!((word)back_ptr & FLAG_MANY)) {
359 if (is_in_progress(p)) return 0; /* DFS back edge, i.e. we followed */
360 /* an edge to an object already */
361 /* on our stack: ignore */
363 result = backwards_height(back_ptr)+1;
367 be = (back_edges *)((word)back_ptr & ~FLAG_MANY);
368 if (be -> height >= 0 && be -> height_gc_no == GC_gc_no)
370 /* Ignore back edges in DFS */
371 if (be -> height == HEIGHT_IN_PROGRESS) return 0;
372 result = (be -> height > 0? be -> height : 1);
373 be -> height = HEIGHT_IN_PROGRESS;
374 FOR_EACH_PRED(q, p, {
376 if (GC_is_marked(q) && !(FLAG_MANY & (word)GET_OH_BG_PTR(p))) {
378 GC_log_printf("Found bogus pointer from 0x%lx to 0x%lx\n", q, p);
379 /* Reachable object "points to" unreachable one. */
380 /* Could be caused by our lax treatment of GC descriptors. */
383 this_height = backwards_height(q);
385 if (this_height >= result) result = this_height + 1;
387 be -> height = result;
388 be -> height_gc_no = GC_gc_no;
393 ptr_t GC_deepest_obj;
395 /* Compute the maximum height of every unreachable predecessor p of a */
396 /* reachable object. Arrange to save the heights of all such objects p */
397 /* so that they can be used in calculating the height of objects in the */
399 /* Set GC_max_height to be the maximum height we encounter, and */
400 /* GC_deepest_obj to be the corresponding object. */
401 static void update_max_height(ptr_t p, size_t n_bytes, word gc_descr)
403 if (GC_is_marked(p) && GC_HAS_DEBUG_INFO(p)) {
406 ptr_t p_deepest_obj = 0;
410 /* If we remembered a height last time, use it as a minimum. */
411 /* It may have increased due to newly unreachable chains pointing */
412 /* to p, but it can't have decreased. */
413 back_ptr = GET_OH_BG_PTR(p);
414 if (0 != back_ptr && ((word)back_ptr & FLAG_MANY)) {
415 be = (back_edges *)((word)back_ptr & ~FLAG_MANY);
416 if (be -> height != HEIGHT_UNKNOWN) p_height = be -> height;
418 FOR_EACH_PRED(q, p, {
419 if (!GC_is_marked(q) && GC_HAS_DEBUG_INFO(q)) {
422 q_height = backwards_height(q);
423 if (q_height > p_height) {
430 /* Remember the height for next time. */
433 back_ptr = GET_OH_BG_PTR(p);
434 be = (back_edges *)((word)back_ptr & ~FLAG_MANY);
436 be -> flags |= RETAIN;
437 be -> height = p_height;
438 be -> height_gc_no = GC_gc_no;
440 if (p_height > GC_max_height) {
441 GC_max_height = p_height;
442 GC_deepest_obj = p_deepest_obj;
447 word GC_max_max_height = 0;
449 void GC_traverse_back_graph(void)
452 GC_apply_to_each_object(update_max_height);
453 if (0 != GC_deepest_obj)
454 GC_set_mark_bit(GC_deepest_obj); /* Keep it until we can print it. */
457 void GC_print_back_graph_stats(void)
459 GC_printf("Maximum backwards height of reachable objects at GC %lu is %ld\n",
460 (unsigned long) GC_gc_no, (unsigned long)GC_max_height);
461 if (GC_max_height > GC_max_max_height) {
462 GC_max_max_height = GC_max_height;
463 GC_printf("The following unreachable object is last in a longest chain "
464 "of unreachable objects:\n");
465 GC_print_heap_obj(GC_deepest_obj);
467 if (GC_print_stats) {
468 GC_log_printf("Needed max total of %ld back-edge structs\n",
469 GC_n_back_edge_structs);
471 GC_apply_to_each_object(reset_back_edge);
475 #endif /* MAKE_BACK_GRAPH */