56 shenandoah_assert_not_in_cset_except(nullptr, obj, ShenandoahHeap::heap()->cancelled_gc());
57
58 Klass* klass = obj->klass();
59
60 // Are we in weak subgraph scan?
61 bool weak = task->is_weak();
62 cl->set_weak(weak);
63
64 if (task->is_not_chunked()) {
65 // Dispatch based on object type. The case order does not seem to affect performance,
66 // so it matches the enum order for consistency.
67 switch (klass->kind()) {
68 case Klass::InstanceKlassKind: {
69 // Regular instance.
70 if (STRING_DEDUP && (klass == vmClasses::String_klass())) {
71 dedup_string(obj, req);
72 }
73 InstanceKlass::cast(klass)->oop_oop_iterate<OT>(obj, cl);
74 break;
75 }
76 case Klass::InstanceRefKlassKind: {
77 // (Weak) reference instance.
78 InstanceRefKlass::cast(klass)->oop_oop_iterate<OT>(obj, cl);
79 break;
80 }
81 case Klass::InstanceMirrorKlassKind:
82 case Klass::InstanceClassLoaderKlassKind: {
83 // Remaining rare classes, dispatch generically.
84 obj->oop_iterate(cl);
85 break;
86 }
87 case Klass::InstanceStackChunkKlassKind: {
88 // Stack chunk. Loom doesn't support mixing of weak marking and strong marking
89 // of stack chunks, upgrade to strong right away.
90 cl->set_weak(false);
91 InstanceStackChunkKlass::cast(klass)->oop_oop_iterate<OT>(obj, cl);
92 break;
93 }
94 case Klass::TypeArrayKlassKind: {
95 // Primitive array. Do nothing, no oops there. We use the same
96 // performance tweak TypeArrayKlass::oop_oop_iterate_impl is using:
97 // We skip iterating over the klass pointer since we know that
98 // Universe::TypeArrayKlass never moves.
99 break;
100 }
101 case Klass::ObjArrayKlassKind: {
102 // Object array and no chunk is set. Must be the first
103 // time we visit it, start the chunked processing.
104 do_chunked_array_start<T, OT>(q, cl, obj, klass, weak);
105 break;
106 }
107 default: {
108 fatal("Unknown klass kind: %d", klass->kind());
109 }
110 }
111 // Count liveness the last: push the outstanding work to the queues first
112 // Avoid double-counting objects that are visited twice due to upgrade
113 // from final- to strong mark.
114 if (task->count_liveness()) {
115 count_liveness<GENERATION>(live_data, obj, klass, worker_id);
116 }
117 } else {
118 // Object array chunk. Process it.
119 do_chunked_array<T, OT>(q, cl, obj, klass, task->chunk(), task->pow(), weak);
120 }
121 }
122
123 inline void ShenandoahMark::dedup_string(oop obj, StringDedup::Requests* const req) {
124 assert(req != nullptr, "Should be available if dedup is enabled");
125
126 // Skip if already requested or dedup is forbidden.
127 // The overwhelming majority of Strings would be filtered here.
128 // These bits are also sticky, so older Strings would be filtered here too.
129 if (java_lang_String::deduplication_requested_or_forbidden(obj)) {
130 return;
131 }
132
133 // Accept deduplication request.
134 if (!java_lang_String::test_and_set_deduplication_requested(obj)) {
135 req->add(obj);
136 }
137 }
138
165 } else {
166 // still good, remember in locals
167 live_data[region_idx] = (ShenandoahLiveData) new_val;
168 }
169 } else {
170 shenandoah_assert_in_correct_region(nullptr, obj);
171 size_t num_regions = ShenandoahHeapRegion::required_regions(size * HeapWordSize);
172
173 assert(region->is_affiliated(), "Do not count live data within FREE Humongous Start Region %zu", region_idx);
174 for (size_t i = region_idx; i < region_idx + num_regions; i++) {
175 ShenandoahHeapRegion* chain_reg = heap->get_region(i);
176 assert(chain_reg->is_humongous(), "Expecting a humongous region");
177 assert(chain_reg->is_affiliated(), "Do not count live data within FREE Humongous Continuation Region %zu", i);
178 chain_reg->increase_live_data_gc_words(chain_reg->used() >> LogHeapWordSize);
179 }
180 }
181 }
182
183 template <class T, class OT>
184 void ShenandoahMark::do_chunked_array_start(ShenandoahObjToScanQueue* q, T* cl, oop obj, Klass* klass, bool weak) {
185 assert(obj->is_objArray(), "expect object array");
186 objArrayOop array = objArrayOop(obj);
187 int len = array->length();
188
189 // Mark objArray klass metadata
190 if (Devirtualizer::do_metadata(cl)) {
191 Devirtualizer::do_klass(cl, klass);
192 }
193
194 if (len <= (int) ObjArrayMarkingStride*2) {
195 // A few slices only, process directly
196 ObjArrayKlass::cast(klass)->oop_oop_iterate_elements_range<OT>(array, cl, 0, len);
197 } else {
198 int bits = log2i_graceful(len);
199 // Compensate for non-power-of-two arrays, cover the array in excess:
200 if (len != (1 << bits)) bits++;
201
202 // Only allow full chunks on the queue. This frees do_chunked_array() from checking from/to
203 // boundaries against array->length(), touching the array header on every chunk.
204 //
205 // To do this, we cut the prefix in full-sized chunks, and submit them on the queue.
206 // If the array is not divided in chunk sizes, then there would be an irregular tail,
207 // which we will process separately.
208
209 int last_idx = 0;
210
211 int chunk = 1;
212 int pow = bits;
213
214 // Handle overflow
215 if (pow >= 31) {
216 assert (pow == 31, "sanity");
225 // successful right boundary to figure out the irregular tail.
226 while ((1 << pow) > (int)ObjArrayMarkingStride &&
227 (chunk*2 < ShenandoahMarkTask::chunk_size())) {
228 pow--;
229 int left_chunk = chunk*2 - 1;
230 int right_chunk = chunk*2;
231 int left_chunk_end = left_chunk * (1 << pow);
232 if (left_chunk_end < len) {
233 bool pushed = q->push(ShenandoahMarkTask(array, true, weak, left_chunk, pow));
234 assert(pushed, "overflow queue should always succeed pushing");
235 chunk = right_chunk;
236 last_idx = left_chunk_end;
237 } else {
238 chunk = left_chunk;
239 }
240 }
241
242 // Process the irregular tail, if present
243 int from = last_idx;
244 if (from < len) {
245 ObjArrayKlass::cast(klass)->oop_oop_iterate_elements_range<OT>(array, cl, from, len);
246 }
247 }
248 }
249
250 template <class T, class OT>
251 void ShenandoahMark::do_chunked_array(ShenandoahObjToScanQueue* q, T* cl, oop obj, Klass* klass, int chunk, int pow, bool weak) {
252 assert(obj->is_objArray(), "expect object array");
253 objArrayOop array = objArrayOop(obj);
254
255 // Split out tasks, as suggested in ShenandoahMarkTask docs. Avoid pushing tasks that
256 // are known to start beyond the array.
257 while ((1 << pow) > (int)ObjArrayMarkingStride && (chunk*2 < ShenandoahMarkTask::chunk_size())) {
258 pow--;
259 chunk *= 2;
260 bool pushed = q->push(ShenandoahMarkTask(array, true, weak, chunk - 1, pow));
261 assert(pushed, "overflow queue should always succeed pushing");
262 }
263
264 int chunk_size = 1 << pow;
265
266 int from = (chunk - 1) * chunk_size;
267 int to = chunk * chunk_size;
268
269 #ifdef ASSERT
270 int len = array->length();
271 assert (0 <= from && from < len, "from is sane: %d/%d", from, len);
272 assert (0 < to && to <= len, "to is sane: %d/%d", to, len);
273 #endif
274
275 ObjArrayKlass::cast(klass)->oop_oop_iterate_elements_range<OT>(array, cl, from, to);
276 }
277
278 template <ShenandoahGenerationType GENERATION>
279 class ShenandoahSATBBufferClosure : public SATBBufferClosure {
280 private:
281 ShenandoahObjToScanQueue* _queue;
282 ShenandoahObjToScanQueue* _old_queue;
283 ShenandoahHeap* _heap;
284 ShenandoahMarkingContext* const _mark_context;
285 public:
286 ShenandoahSATBBufferClosure(ShenandoahObjToScanQueue* q, ShenandoahObjToScanQueue* old_q) :
287 _queue(q),
288 _old_queue(old_q),
289 _heap(ShenandoahHeap::heap()),
290 _mark_context(_heap->marking_context())
291 {
292 }
293
294 void do_buffer(void **buffer, size_t size) {
295 assert(size == 0 || !_heap->has_forwarded_objects() || _heap->is_concurrent_old_mark_in_progress(), "Forwarded objects are not expected here");
|
56 shenandoah_assert_not_in_cset_except(nullptr, obj, ShenandoahHeap::heap()->cancelled_gc());
57
58 Klass* klass = obj->klass();
59
60 // Are we in weak subgraph scan?
61 bool weak = task->is_weak();
62 cl->set_weak(weak);
63
64 if (task->is_not_chunked()) {
65 // Dispatch based on object type. The case order does not seem to affect performance,
66 // so it matches the enum order for consistency.
67 switch (klass->kind()) {
68 case Klass::InstanceKlassKind: {
69 // Regular instance.
70 if (STRING_DEDUP && (klass == vmClasses::String_klass())) {
71 dedup_string(obj, req);
72 }
73 InstanceKlass::cast(klass)->oop_oop_iterate<OT>(obj, cl);
74 break;
75 }
76 case Klass::InlineKlassKind: {
77 // Inline instance.
78 InlineKlass::cast(klass)->oop_oop_iterate<OT>(obj, cl);
79 break;
80 }
81 case Klass::InstanceRefKlassKind: {
82 // (Weak) reference instance.
83 InstanceRefKlass::cast(klass)->oop_oop_iterate<OT>(obj, cl);
84 break;
85 }
86 case Klass::InstanceMirrorKlassKind:
87 case Klass::InstanceClassLoaderKlassKind: {
88 // Remaining rare classes, dispatch generically.
89 obj->oop_iterate(cl);
90 break;
91 }
92 case Klass::InstanceStackChunkKlassKind: {
93 // Stack chunk. Loom doesn't support mixing of weak marking and strong marking
94 // of stack chunks, upgrade to strong right away.
95 cl->set_weak(false);
96 InstanceStackChunkKlass::cast(klass)->oop_oop_iterate<OT>(obj, cl);
97 break;
98 }
99 case Klass::TypeArrayKlassKind: {
100 // Primitive array. Do nothing, no oops there. We use the same
101 // performance tweak TypeArrayKlass::oop_oop_iterate_impl is using:
102 // We skip iterating over the klass pointer since we know that
103 // Universe::TypeArrayKlass never moves.
104 break;
105 }
106 case Klass::ObjArrayKlassKind: {
107 fatal("Unexpected unrefined object array klass");
108 }
109 case Klass::RefArrayKlassKind: {
110 // Reference array and no chunk is set. Must be the first
111 // time we visit it, start the chunked processing.
112 do_chunked_array_start<T, OT>(q, cl, obj, klass, weak);
113 break;
114 }
115 case Klass::FlatArrayKlassKind: {
116 // Flat array, all elements are embedded.
117 FlatArrayKlass::cast(klass)->oop_oop_iterate<OT>(obj, cl);
118 break;
119 }
120 default: {
121 fatal("Unknown klass kind: %d", klass->kind());
122 }
123 }
124 // Count liveness the last: push the outstanding work to the queues first
125 // Avoid double-counting objects that are visited twice due to upgrade
126 // from final- to strong mark.
127 if (task->count_liveness()) {
128 count_liveness<GENERATION>(live_data, obj, klass, worker_id);
129 }
130 } else {
131 // Reference array chunk. Process it.
132 do_chunked_array<T, OT>(q, cl, obj, klass, task->chunk(), task->pow(), weak);
133 }
134 }
135
136 inline void ShenandoahMark::dedup_string(oop obj, StringDedup::Requests* const req) {
137 assert(req != nullptr, "Should be available if dedup is enabled");
138
139 // Skip if already requested or dedup is forbidden.
140 // The overwhelming majority of Strings would be filtered here.
141 // These bits are also sticky, so older Strings would be filtered here too.
142 if (java_lang_String::deduplication_requested_or_forbidden(obj)) {
143 return;
144 }
145
146 // Accept deduplication request.
147 if (!java_lang_String::test_and_set_deduplication_requested(obj)) {
148 req->add(obj);
149 }
150 }
151
178 } else {
179 // still good, remember in locals
180 live_data[region_idx] = (ShenandoahLiveData) new_val;
181 }
182 } else {
183 shenandoah_assert_in_correct_region(nullptr, obj);
184 size_t num_regions = ShenandoahHeapRegion::required_regions(size * HeapWordSize);
185
186 assert(region->is_affiliated(), "Do not count live data within FREE Humongous Start Region %zu", region_idx);
187 for (size_t i = region_idx; i < region_idx + num_regions; i++) {
188 ShenandoahHeapRegion* chain_reg = heap->get_region(i);
189 assert(chain_reg->is_humongous(), "Expecting a humongous region");
190 assert(chain_reg->is_affiliated(), "Do not count live data within FREE Humongous Continuation Region %zu", i);
191 chain_reg->increase_live_data_gc_words(chain_reg->used() >> LogHeapWordSize);
192 }
193 }
194 }
195
196 template <class T, class OT>
197 void ShenandoahMark::do_chunked_array_start(ShenandoahObjToScanQueue* q, T* cl, oop obj, Klass* klass, bool weak) {
198 assert(obj->is_refArray(), "expect ref array");
199 refArrayOop array = refArrayOop(obj);
200 int len = array->length();
201
202 // Mark reference array klass metadata
203 if (Devirtualizer::do_metadata(cl)) {
204 Devirtualizer::do_klass(cl, klass);
205 }
206
207 if (len <= (int) ObjArrayMarkingStride*2) {
208 // A few slices only, process directly
209 RefArrayKlass::cast(klass)->oop_oop_iterate_elements_range<OT>(array, cl, 0, len);
210 } else {
211 int bits = log2i_graceful(len);
212 // Compensate for non-power-of-two arrays, cover the array in excess:
213 if (len != (1 << bits)) bits++;
214
215 // Only allow full chunks on the queue. This frees do_chunked_array() from checking from/to
216 // boundaries against array->length(), touching the array header on every chunk.
217 //
218 // To do this, we cut the prefix in full-sized chunks, and submit them on the queue.
219 // If the array is not divided in chunk sizes, then there would be an irregular tail,
220 // which we will process separately.
221
222 int last_idx = 0;
223
224 int chunk = 1;
225 int pow = bits;
226
227 // Handle overflow
228 if (pow >= 31) {
229 assert (pow == 31, "sanity");
238 // successful right boundary to figure out the irregular tail.
239 while ((1 << pow) > (int)ObjArrayMarkingStride &&
240 (chunk*2 < ShenandoahMarkTask::chunk_size())) {
241 pow--;
242 int left_chunk = chunk*2 - 1;
243 int right_chunk = chunk*2;
244 int left_chunk_end = left_chunk * (1 << pow);
245 if (left_chunk_end < len) {
246 bool pushed = q->push(ShenandoahMarkTask(array, true, weak, left_chunk, pow));
247 assert(pushed, "overflow queue should always succeed pushing");
248 chunk = right_chunk;
249 last_idx = left_chunk_end;
250 } else {
251 chunk = left_chunk;
252 }
253 }
254
255 // Process the irregular tail, if present
256 int from = last_idx;
257 if (from < len) {
258 RefArrayKlass::cast(klass)->oop_oop_iterate_elements_range<OT>(array, cl, from, len);
259 }
260 }
261 }
262
263 template <class T, class OT>
264 void ShenandoahMark::do_chunked_array(ShenandoahObjToScanQueue* q, T* cl, oop obj, Klass* klass, int chunk, int pow, bool weak) {
265 assert(obj->is_refArray(), "expect ref array");
266 refArrayOop array = refArrayOop(obj);
267
268 // Split out tasks, as suggested in ShenandoahMarkTask docs. Avoid pushing tasks that
269 // are known to start beyond the array.
270 while ((1 << pow) > (int)ObjArrayMarkingStride && (chunk*2 < ShenandoahMarkTask::chunk_size())) {
271 pow--;
272 chunk *= 2;
273 bool pushed = q->push(ShenandoahMarkTask(array, true, weak, chunk - 1, pow));
274 assert(pushed, "overflow queue should always succeed pushing");
275 }
276
277 int chunk_size = 1 << pow;
278
279 int from = (chunk - 1) * chunk_size;
280 int to = chunk * chunk_size;
281
282 #ifdef ASSERT
283 int len = array->length();
284 assert (0 <= from && from < len, "from is sane: %d/%d", from, len);
285 assert (0 < to && to <= len, "to is sane: %d/%d", to, len);
286 #endif
287
288 RefArrayKlass::cast(klass)->oop_oop_iterate_elements_range<OT>(array, cl, from, to);
289 }
290
291 template <ShenandoahGenerationType GENERATION>
292 class ShenandoahSATBBufferClosure : public SATBBufferClosure {
293 private:
294 ShenandoahObjToScanQueue* _queue;
295 ShenandoahObjToScanQueue* _old_queue;
296 ShenandoahHeap* _heap;
297 ShenandoahMarkingContext* const _mark_context;
298 public:
299 ShenandoahSATBBufferClosure(ShenandoahObjToScanQueue* q, ShenandoahObjToScanQueue* old_q) :
300 _queue(q),
301 _old_queue(old_q),
302 _heap(ShenandoahHeap::heap()),
303 _mark_context(_heap->marking_context())
304 {
305 }
306
307 void do_buffer(void **buffer, size_t size) {
308 assert(size == 0 || !_heap->has_forwarded_objects() || _heap->is_concurrent_old_mark_in_progress(), "Forwarded objects are not expected here");
|