1 /*
2 * Copyright (c) 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/g1/c2/g1BarrierSetC2.hpp"
27 #include "gc/g1/g1BarrierSet.hpp"
28 #include "gc/g1/g1BarrierSetRuntime.hpp"
29 #include "gc/g1/g1CardTable.hpp"
30 #include "gc/g1/g1ThreadLocalData.hpp"
31 #include "gc/g1/heapRegion.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/compile.hpp"
34 #include "opto/escape.hpp"
35 #include "opto/graphKit.hpp"
36 #include "opto/idealKit.hpp"
37 #include "opto/macro.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/type.hpp"
40 #include "utilities/macros.hpp"
41
42 const TypeFunc *G1BarrierSetC2::write_ref_field_pre_entry_Type() {
43 const Type **fields = TypeTuple::fields(2);
44 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
45 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
46 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
47
48 // create result type (range)
49 fields = TypeTuple::fields(0);
50 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
51
52 return TypeFunc::make(domain, range);
53 }
54
55 const TypeFunc *G1BarrierSetC2::write_ref_field_post_entry_Type() {
56 const Type **fields = TypeTuple::fields(2);
57 fields[TypeFunc::Parms+0] = TypeRawPtr::NOTNULL; // Card addr
58 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
59 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
60
61 // create result type (range)
62 fields = TypeTuple::fields(0);
63 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms, fields);
64
65 return TypeFunc::make(domain, range);
66 }
67
68 const TypeFunc *G1BarrierSetC2::write_ref_stats_Type() {
69 const Type **fields = TypeTuple::fields(2);
70 fields[TypeFunc::Parms+0] = TypeRawPtr::NOTNULL; // HeapWord* dst
71 fields[TypeFunc::Parms+1] = TypeInstPtr::NOTNULL; // oopDesc* val
72 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
73
74 // create result type (range)
75 fields = TypeTuple::fields(0);
76 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms, fields);
77
78 return TypeFunc::make(domain, range);
79 }
80
81 #define __ ideal.
82 /*
83 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
84 * required by SATB to make sure all objects live at the start of the
85 * marking are kept alive, all reference updates need to any previous
86 * reference stored before writing.
87 *
88 * If the previous value is NULL there is no need to save the old value.
89 * References that are NULL are filtered during runtime by the barrier
90 * code to avoid unnecessary queuing.
91 *
92 * However in the case of newly allocated objects it might be possible to
93 * prove that the reference about to be overwritten is NULL during compile
94 * time and avoid adding the barrier code completely.
95 *
96 * The compiler needs to determine that the object in which a field is about
97 * to be written is newly allocated, and that no prior store to the same field
98 * has happened since the allocation.
99 *
100 * Returns true if the pre-barrier can be removed
101 */
102 bool G1BarrierSetC2::g1_can_remove_pre_barrier(GraphKit* kit,
103 PhaseTransform* phase,
104 Node* adr,
105 BasicType bt,
106 uint adr_idx) const {
107 intptr_t offset = 0;
108 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
109 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
110
111 if (offset == Type::OffsetBot) {
112 return false; // cannot unalias unless there are precise offsets
113 }
114
115 if (alloc == NULL) {
116 return false; // No allocation found
117 }
118
119 intptr_t size_in_bytes = type2aelembytes(bt);
120
121 Node* mem = kit->memory(adr_idx); // start searching here...
122
123 for (int cnt = 0; cnt < 50; cnt++) {
124
125 if (mem->is_Store()) {
126
127 Node* st_adr = mem->in(MemNode::Address);
128 intptr_t st_offset = 0;
129 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
130
131 if (st_base == NULL) {
132 break; // inscrutable pointer
133 }
134
135 // Break we have found a store with same base and offset as ours so break
136 if (st_base == base && st_offset == offset) {
137 break;
138 }
139
140 if (st_offset != offset && st_offset != Type::OffsetBot) {
141 const int MAX_STORE = BytesPerLong;
142 if (st_offset >= offset + size_in_bytes ||
143 st_offset <= offset - MAX_STORE ||
144 st_offset <= offset - mem->as_Store()->memory_size()) {
145 // Success: The offsets are provably independent.
146 // (You may ask, why not just test st_offset != offset and be done?
147 // The answer is that stores of different sizes can co-exist
148 // in the same sequence of RawMem effects. We sometimes initialize
149 // a whole 'tile' of array elements with a single jint or jlong.)
150 mem = mem->in(MemNode::Memory);
151 continue; // advance through independent store memory
152 }
153 }
154
155 if (st_base != base
156 && MemNode::detect_ptr_independence(base, alloc, st_base,
157 AllocateNode::Ideal_allocation(st_base, phase),
158 phase)) {
159 // Success: The bases are provably independent.
160 mem = mem->in(MemNode::Memory);
161 continue; // advance through independent store memory
162 }
163 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
164
165 InitializeNode* st_init = mem->in(0)->as_Initialize();
166 AllocateNode* st_alloc = st_init->allocation();
167
168 // Make sure that we are looking at the same allocation site.
169 // The alloc variable is guaranteed to not be null here from earlier check.
170 if (alloc == st_alloc) {
171 // Check that the initialization is storing NULL so that no previous store
172 // has been moved up and directly write a reference
173 Node* captured_store = st_init->find_captured_store(offset,
174 type2aelembytes(T_OBJECT),
175 phase);
176 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
177 return true;
178 }
179 }
180 }
181
182 // Unless there is an explicit 'continue', we must bail out here,
183 // because 'mem' is an inscrutable memory state (e.g., a call).
184 break;
185 }
186
187 return false;
188 }
189
190 // G1 pre/post barriers
191 void G1BarrierSetC2::pre_barrier(GraphKit* kit,
192 bool do_load,
193 Node* ctl,
194 Node* obj,
195 Node* adr,
196 uint alias_idx,
197 Node* val,
198 const TypeOopPtr* val_type,
199 Node* pre_val,
200 BasicType bt) const {
201 // Some sanity checks
202 // Note: val is unused in this routine.
203
204 if (do_load) {
205 // We need to generate the load of the previous value
206 assert(obj != NULL, "must have a base");
207 assert(adr != NULL, "where are loading from?");
208 assert(pre_val == NULL, "loaded already?");
209 assert(val_type != NULL, "need a type");
210
211 if (use_ReduceInitialCardMarks()
212 && g1_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
213 return;
214 }
215
216 } else {
217 // In this case both val_type and alias_idx are unused.
218 assert(pre_val != NULL, "must be loaded already");
219 // Nothing to be done if pre_val is null.
220 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
221 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
222 }
223 assert(bt == T_OBJECT, "or we shouldn't be here");
224
225 IdealKit ideal(kit, true);
226
227 Node* tls = __ thread(); // ThreadLocalStorage
228
229 Node* no_base = __ top();
230 Node* zero = __ ConI(0);
231 Node* zeroX = __ ConX(0);
232
233 float likely = PROB_LIKELY(0.999);
234 float unlikely = PROB_UNLIKELY(0.999);
235
236 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
237 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
238
239 // Offsets into the thread
240 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
241 const int index_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset());
242 const int buffer_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset());
243
244 // Now the actual pointers into the thread
245 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
246 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
247 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
248
249 // Now some of the values
250 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
251
252 // if (!marking)
253 __ if_then(marking, BoolTest::ne, zero, unlikely); {
254 BasicType index_bt = TypeX_X->basic_type();
255 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
256 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
257
258 if (do_load) {
259 // load original value
260 // alias_idx correct??
261 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
262 }
263
264 // if (pre_val != NULL)
265 __ if_then(pre_val, BoolTest::ne, kit->null()); {
266 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
267
268 // is the queue for this thread full?
269 __ if_then(index, BoolTest::ne, zeroX, likely); {
270
271 // decrement the index
272 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
273
274 // Now get the buffer location we will log the previous value into and store it
275 Node *log_addr = __ AddP(no_base, buffer, next_index);
276 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
277 // update the index
278 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
279
280 } __ else_(); {
281
282 // logging buffer is full, call the runtime
283 const TypeFunc *tf = write_ref_field_pre_entry_Type();
284 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), "write_ref_field_pre_entry", pre_val, tls);
285 } __ end_if(); // (!index)
286 } __ end_if(); // (pre_val != NULL)
287 } __ end_if(); // (!marking)
288
289 // Final sync IdealKit and GraphKit.
290 kit->final_sync(ideal);
291 }
292
293 /*
294 * G1 similar to any GC with a Young Generation requires a way to keep track of
295 * references from Old Generation to Young Generation to make sure all live
296 * objects are found. G1 also requires to keep track of object references
297 * between different regions to enable evacuation of old regions, which is done
298 * as part of mixed collections. References are tracked in remembered sets and
299 * is continuously updated as reference are written to with the help of the
300 * post-barrier.
301 *
302 * To reduce the number of updates to the remembered set the post-barrier
303 * filters updates to fields in objects located in the Young Generation,
304 * the same region as the reference, when the NULL is being written or
305 * if the card is already marked as dirty by an earlier write.
306 *
307 * Under certain circumstances it is possible to avoid generating the
308 * post-barrier completely if it is possible during compile time to prove
309 * the object is newly allocated and that no safepoint exists between the
310 * allocation and the store.
311 *
312 * In the case of slow allocation the allocation code must handle the barrier
313 * as part of the allocation in the case the allocated object is not located
314 * in the nursery, this would happen for humongous objects. This is similar to
315 * how CMS is required to handle this case, see the comments for the method
316 * CollectedHeap::new_deferred_store_barrier and OptoRuntime::new_deferred_store_barrier.
317 * A deferred card mark is required for these objects and handled in the above
318 * mentioned methods.
319 *
320 * Returns true if the post barrier can be removed
321 */
322 bool G1BarrierSetC2::g1_can_remove_post_barrier(GraphKit* kit,
323 PhaseTransform* phase, Node* store,
324 Node* adr) const {
325 intptr_t offset = 0;
326 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
327 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
328
329 if (offset == Type::OffsetBot) {
330 return false; // cannot unalias unless there are precise offsets
331 }
332
333 if (alloc == NULL) {
334 return false; // No allocation found
335 }
336
337 // Start search from Store node
338 Node* mem = store->in(MemNode::Control);
339 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
340
341 InitializeNode* st_init = mem->in(0)->as_Initialize();
342 AllocateNode* st_alloc = st_init->allocation();
343
344 // Make sure we are looking at the same allocation
345 if (alloc == st_alloc) {
346 return true;
347 }
348 }
349
350 return false;
351 }
352
353 //
354 // Update the card table and add card address to the queue
355 //
356 void G1BarrierSetC2::g1_mark_card(GraphKit* kit,
357 IdealKit& ideal,
358 Node* card_adr,
359 Node* oop_store,
360 uint oop_alias_idx,
361 Node* index,
362 Node* index_adr,
363 Node* buffer,
364 const TypeFunc* tf) const {
365 Node* zero = __ ConI(0);
366 Node* zeroX = __ ConX(0);
367 Node* no_base = __ top();
368 BasicType card_bt = T_BYTE;
369 // Smash zero into card. MUST BE ORDERED WRT TO STORE
370 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
371
372 // Now do the queue work
373 __ if_then(index, BoolTest::ne, zeroX); {
374
375 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
376 Node* log_addr = __ AddP(no_base, buffer, next_index);
377
378 // Order, see storeCM.
379 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
380 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
381
382 } __ else_(); {
383 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_post_entry), "write_ref_field_post_entry", card_adr, __ thread());
384 } __ end_if();
385
386 }
387
388 void G1BarrierSetC2::post_barrier(GraphKit* kit,
389 Node* ctl,
390 Node* oop_store,
391 Node* obj,
392 Node* adr,
393 uint alias_idx,
394 Node* val,
395 BasicType bt,
396 bool use_precise) const {
397 // If we are writing a NULL then we need no post barrier
398
399 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
400 // Must be NULL
401 const Type* t = val->bottom_type();
402 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
403 // No post barrier if writing NULLx
404 return;
405 }
406
407 if (use_ReduceInitialCardMarks() && obj == kit->just_allocated_object(kit->control())) {
408 // We can skip marks on a freshly-allocated object in Eden.
409 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
410 // That routine informs GC to take appropriate compensating steps,
411 // upon a slow-path allocation, so as to make this card-mark
412 // elision safe.
413 return;
414 }
415
416 if (use_ReduceInitialCardMarks()
417 && g1_can_remove_post_barrier(kit, &kit->gvn(), oop_store, adr)) {
418 return;
419 }
420
421 if (!use_precise) {
422 // All card marks for a (non-array) instance are in one place:
423 adr = obj;
424 }
425 // (Else it's an array (or unknown), and we want more precise card marks.)
426 assert(adr != NULL, "");
427
428 IdealKit ideal(kit, true);
429
430 Node* tls = __ thread(); // ThreadLocalStorage
431
432 Node* no_base = __ top();
433 float unlikely = PROB_UNLIKELY(0.999);
434 Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val());
435 Node* dirty_card = __ ConI((jint)G1CardTable::dirty_card_val());
436 Node* zeroX = __ ConX(0);
437
438 const TypeFunc *tf = write_ref_field_post_entry_Type();
439
440 // Offsets into the thread
441 const int index_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset());
442 const int buffer_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset());
443
444 // Pointers into the thread
445
446 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
447 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
448
449 // Now some values
450 // Use ctrl to avoid hoisting these values past a safepoint, which could
451 // potentially reset these fields in the JavaThread.
452 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
453 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
454
455 // Convert the store obj pointer to an int prior to doing math on it
456 // Must use ctrl to prevent "integerized oop" existing across safepoint
457 Node* cast = __ CastPX(__ ctrl(), adr);
458
459 // Divide pointer by card size
460 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
461
462 // Combine card table base and card offset
463 Node* card_adr = __ AddP(no_base, byte_map_base_node(kit), card_offset );
464
465 // If we know the value being stored does it cross regions?
466
467 if (val != NULL) {
468 if (UsePerfData && G1WriteBarrierStats) {
469 const TypeFunc *stats_tf = write_ref_stats_Type();
470 __ make_leaf_call(stats_tf, CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_stats), "write_ref_stats", adr, val);
471 }
472 // Does the store cause us to cross regions?
473
474 // Should be able to do an unsigned compare of region_size instead of
475 // and extra shift. Do we have an unsigned compare??
476 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
477 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
478
479 // if (xor_res == 0) same region so skip
480 __ if_then(xor_res, BoolTest::ne, zeroX); {
481
482 // No barrier if we are storing a NULL
483 __ if_then(val, BoolTest::ne, kit->null(), unlikely); {
484
485 // Ok must mark the card if not already dirty
486
487 // load the original value of the card
488 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
489
490 __ if_then(card_val, BoolTest::ne, young_card); {
491 kit->sync_kit(ideal);
492 kit->insert_store_load_for_barrier();
493 __ sync_kit(kit);
494
495 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
496 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
497 g1_mark_card(kit, ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
498 } __ end_if();
499 } __ end_if();
500 } __ end_if();
501 } __ end_if();
502 } else {
503 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
504 // We don't need a barrier here if the destination is a newly allocated object
505 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
506 // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()).
507 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
508 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
509 __ if_then(card_val, BoolTest::ne, young_card); {
510 g1_mark_card(kit, ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
511 } __ end_if();
512 }
513
514 // Final sync IdealKit and GraphKit.
515 kit->final_sync(ideal);
516 }
517
518 // Helper that guards and inserts a pre-barrier.
519 void G1BarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
520 Node* pre_val, bool need_mem_bar) const {
521 // We could be accessing the referent field of a reference object. If so, when G1
522 // is enabled, we need to log the value in the referent field in an SATB buffer.
523 // This routine performs some compile time filters and generates suitable
524 // runtime filters that guard the pre-barrier code.
525 // Also add memory barrier for non volatile load from the referent field
526 // to prevent commoning of loads across safepoint.
527
528 // Some compile time checks.
529
530 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
531 const TypeX* otype = offset->find_intptr_t_type();
532 if (otype != NULL && otype->is_con() &&
533 otype->get_con() != java_lang_ref_Reference::referent_offset) {
534 // Constant offset but not the reference_offset so just return
535 return;
536 }
537
538 // We only need to generate the runtime guards for instances.
539 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
540 if (btype != NULL) {
541 if (btype->isa_aryptr()) {
542 // Array type so nothing to do
543 return;
544 }
545
546 const TypeInstPtr* itype = btype->isa_instptr();
547 if (itype != NULL) {
548 // Can the klass of base_oop be statically determined to be
549 // _not_ a sub-class of Reference and _not_ Object?
550 ciKlass* klass = itype->klass();
551 if ( klass->is_loaded() &&
552 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
553 !kit->env()->Object_klass()->is_subtype_of(klass)) {
554 return;
555 }
556 }
557 }
558
559 // The compile time filters did not reject base_oop/offset so
560 // we need to generate the following runtime filters
561 //
562 // if (offset == java_lang_ref_Reference::_reference_offset) {
563 // if (instance_of(base, java.lang.ref.Reference)) {
564 // pre_barrier(_, pre_val, ...);
565 // }
566 // }
567
568 float likely = PROB_LIKELY( 0.999);
569 float unlikely = PROB_UNLIKELY(0.999);
570
571 IdealKit ideal(kit);
572
573 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
574
575 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
576 // Update graphKit memory and control from IdealKit.
577 kit->sync_kit(ideal);
578
579 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
580 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
581
582 // Update IdealKit memory and control from graphKit.
583 __ sync_kit(kit);
584
585 Node* one = __ ConI(1);
586 // is_instof == 0 if base_oop == NULL
587 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
588
589 // Update graphKit from IdeakKit.
590 kit->sync_kit(ideal);
591
592 // Use the pre-barrier to record the value in the referent field
593 pre_barrier(kit, false /* do_load */,
594 __ ctrl(),
595 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
596 pre_val /* pre_val */,
597 T_OBJECT);
598 if (need_mem_bar) {
599 // Add memory barrier to prevent commoning reads from this field
600 // across safepoint since GC can change its value.
601 kit->insert_mem_bar(Op_MemBarCPUOrder);
602 }
603 // Update IdealKit from graphKit.
604 __ sync_kit(kit);
605
606 } __ end_if(); // _ref_type != ref_none
607 } __ end_if(); // offset == referent_offset
608
609 // Final sync IdealKit and GraphKit.
610 kit->final_sync(ideal);
611 }
612
613 #undef __
614
615 Node* G1BarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
616 DecoratorSet decorators = access.decorators();
617 Node* adr = access.addr().node();
618 Node* obj = access.base();
619
620 bool mismatched = (decorators & C2_MISMATCHED) != 0;
621 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
622 bool in_heap = (decorators & IN_HEAP) != 0;
623 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
624 bool is_unordered = (decorators & MO_UNORDERED) != 0;
625 bool need_cpu_mem_bar = !is_unordered || mismatched || !in_heap;
626
627 Node* top = Compile::current()->top();
628 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
629 Node* load = CardTableBarrierSetC2::load_at_resolved(access, val_type);
630
631 // If we are reading the value of the referent field of a Reference
632 // object (either by using Unsafe directly or through reflection)
633 // then, if G1 is enabled, we need to record the referent in an
634 // SATB log buffer using the pre-barrier mechanism.
635 // Also we need to add memory barrier to prevent commoning reads
636 // from this field across safepoint since GC can change its value.
637 bool need_read_barrier = in_heap && (on_weak ||
638 (unknown && offset != top && obj != top));
639
640 if (!access.is_oop() || !need_read_barrier) {
641 return load;
642 }
643
644 assert(access.is_parse_access(), "entry not supported at optimization time");
645 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
646 GraphKit* kit = parse_access.kit();
647
648 if (on_weak) {
649 // Use the pre-barrier to record the value in the referent field
650 pre_barrier(kit, false /* do_load */,
651 kit->control(),
652 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
653 load /* pre_val */, T_OBJECT);
654 // Add memory barrier to prevent commoning reads from this field
655 // across safepoint since GC can change its value.
656 kit->insert_mem_bar(Op_MemBarCPUOrder);
657 } else if (unknown) {
658 // We do not require a mem bar inside pre_barrier if need_mem_bar
659 // is set: the barriers would be emitted by us.
660 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
661 }
662
663 return load;
664 }
665
666 bool G1BarrierSetC2::is_gc_barrier_node(Node* node) const {
667 if (CardTableBarrierSetC2::is_gc_barrier_node(node)) {
668 return true;
669 }
670 if (node->Opcode() != Op_CallLeaf) {
671 return false;
672 }
673 CallLeafNode *call = node->as_CallLeaf();
674 if (call->_name == NULL) {
675 return false;
676 }
677
678 return strcmp(call->_name, "write_ref_field_pre_entry") == 0 || strcmp(call->_name, "write_ref_field_post_entry") == 0 || strcmp(call->_name, "write_ref_stats") == 0;
679 }
680
681 void G1BarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const {
682 assert(node->Opcode() == Op_CastP2X, "ConvP2XNode required");
683 assert(node->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
684 // It could be only one user, URShift node, in Object.clone() intrinsic
685 // but the new allocation is passed to arraycopy stub and it could not
686 // be scalar replaced. So we don't check the case.
687
688 // An other case of only one user (Xor) is when the value check for NULL
689 // in G1 post barrier is folded after CCP so the code which used URShift
690 // is removed.
691
692 // Take Region node before eliminating post barrier since it also
693 // eliminates CastP2X node when it has only one user.
694 Node* this_region = node->in(0);
695 assert(this_region != NULL, "");
696
697 // Remove G1 post barrier.
698
699 // Search for CastP2X->Xor->URShift->Cmp path which
700 // checks if the store done to a different from the value's region.
701 // And replace Cmp with #0 (false) to collapse G1 post barrier.
702 Node* xorx = node->find_out_with(Op_XorX);
703 if (xorx != NULL) {
704 Node* shift = xorx->unique_out();
705 Node* cmpx = shift->unique_out();
706 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
707 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
708 "missing region check in G1 post barrier");
709 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
710
711 // Remove G1 pre barrier.
712
713 // Search "if (marking != 0)" check and set it to "false".
714 // There is no G1 pre barrier if previous stored value is NULL
715 // (for example, after initialization).
716 if (this_region->is_Region() && this_region->req() == 3) {
717 int ind = 1;
718 if (!this_region->in(ind)->is_IfFalse()) {
719 ind = 2;
720 }
721 if (this_region->in(ind)->is_IfFalse() &&
722 this_region->in(ind)->in(0)->Opcode() == Op_If) {
723 Node* bol = this_region->in(ind)->in(0)->in(1);
724 assert(bol->is_Bool(), "");
725 cmpx = bol->in(1);
726 if (bol->as_Bool()->_test._test == BoolTest::ne &&
727 cmpx->is_Cmp() && cmpx->in(2) == macro->intcon(0) &&
728 cmpx->in(1)->is_Load()) {
729 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
730 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
731 if (adr->is_AddP() && adr->in(AddPNode::Base) == macro->top() &&
732 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
733 adr->in(AddPNode::Offset) == macro->MakeConX(marking_offset)) {
734 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
735 }
736 }
737 }
738 }
739 } else {
740 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
741 // This is a G1 post barrier emitted by the Object.clone() intrinsic.
742 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card
743 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier.
744 Node* shift = node->find_out_with(Op_URShiftX);
745 assert(shift != NULL, "missing G1 post barrier");
746 Node* addp = shift->unique_out();
747 Node* load = addp->find_out_with(Op_LoadB);
748 assert(load != NULL, "missing G1 post barrier");
749 Node* cmpx = load->unique_out();
750 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
751 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
752 "missing card value check in G1 post barrier");
753 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
754 // There is no G1 pre barrier in this case
755 }
756 // Now CastP2X can be removed since it is used only on dead path
757 // which currently still alive until igvn optimize it.
758 assert(node->outcnt() == 0 || node->unique_out()->Opcode() == Op_URShiftX, "");
759 macro->replace_node(node, macro->top());
760 }
761
762 Node* G1BarrierSetC2::step_over_gc_barrier(Node* c) const {
763 if (!use_ReduceInitialCardMarks() &&
764 c != NULL && c->is_Region() && c->req() == 3) {
765 for (uint i = 1; i < c->req(); i++) {
766 if (c->in(i) != NULL && c->in(i)->is_Region() &&
767 c->in(i)->req() == 3) {
768 Node* r = c->in(i);
769 for (uint j = 1; j < r->req(); j++) {
770 if (r->in(j) != NULL && r->in(j)->is_Proj() &&
771 r->in(j)->in(0) != NULL &&
772 r->in(j)->in(0)->Opcode() == Op_CallLeaf &&
773 r->in(j)->in(0)->as_Call()->entry_point() == CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_post_entry)) {
774 Node* call = r->in(j)->in(0);
775 c = c->in(i == 1 ? 2 : 1);
776 if (c != NULL) {
777 c = c->in(0);
778 if (c != NULL) {
779 c = c->in(0);
780 assert(call->in(0) == NULL ||
781 call->in(0)->in(0) == NULL ||
782 call->in(0)->in(0)->in(0) == NULL ||
783 call->in(0)->in(0)->in(0)->in(0) == NULL ||
784 call->in(0)->in(0)->in(0)->in(0)->in(0) == NULL ||
785 c == call->in(0)->in(0)->in(0)->in(0)->in(0), "bad barrier shape");
786 return c;
787 }
788 }
789 }
790 }
791 }
792 }
793 }
794 return c;
795 }
796
797 #ifdef ASSERT
798 void G1BarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
799 if (phase != BarrierSetC2::BeforeCodeGen) {
800 return;
801 }
802 // Verify G1 pre-barriers
803 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
804
805 ResourceArea *area = Thread::current()->resource_area();
806 Unique_Node_List visited(area);
807 Node_List worklist(area);
808 // We're going to walk control flow backwards starting from the Root
809 worklist.push(compile->root());
810 while (worklist.size() > 0) {
811 Node* x = worklist.pop();
812 if (x == NULL || x == compile->top()) continue;
813 if (visited.member(x)) {
814 continue;
815 } else {
816 visited.push(x);
817 }
818
819 if (x->is_Region()) {
820 for (uint i = 1; i < x->req(); i++) {
821 worklist.push(x->in(i));
822 }
823 } else {
824 worklist.push(x->in(0));
825 // We are looking for the pattern:
826 // /->ThreadLocal
827 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
828 // \->ConI(0)
829 // We want to verify that the If and the LoadB have the same control
830 // See GraphKit::g1_write_barrier_pre()
831 if (x->is_If()) {
832 IfNode *iff = x->as_If();
833 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
834 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
835 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
836 && cmp->in(1)->is_Load()) {
837 LoadNode* load = cmp->in(1)->as_Load();
838 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
839 && load->in(2)->in(3)->is_Con()
840 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
841
842 Node* if_ctrl = iff->in(0);
843 Node* load_ctrl = load->in(0);
844
845 if (if_ctrl != load_ctrl) {
846 // Skip possible CProj->NeverBranch in infinite loops
847 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
848 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
849 if_ctrl = if_ctrl->in(0)->in(0);
850 }
851 }
852 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
853 }
854 }
855 }
856 }
857 }
858 }
859 }
860 #endif
861
862 bool G1BarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
863 if (opcode == Op_StoreP) {
864 Node* adr = n->in(MemNode::Address);
865 const Type* adr_type = gvn->type(adr);
866 // Pointer stores in G1 barriers looks like unsafe access.
867 // Ignore such stores to be able scalar replace non-escaping
868 // allocations.
869 if (adr_type->isa_rawptr() && adr->is_AddP()) {
870 Node* base = conn_graph->get_addp_base(adr);
871 if (base->Opcode() == Op_LoadP &&
872 base->in(MemNode::Address)->is_AddP()) {
873 adr = base->in(MemNode::Address);
874 Node* tls = conn_graph->get_addp_base(adr);
875 if (tls->Opcode() == Op_ThreadLocal) {
876 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
877 const int buf_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset());
878 if (offs == buf_offset) {
879 return true; // G1 pre barrier previous oop value store.
880 }
881 if (offs == in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset())) {
882 return true; // G1 post barrier card address store.
883 }
884 }
885 }
886 }
887 }
888 return false;
889 }