1 /*
2 * Copyright (c) 2017, 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 "opto/arraycopynode.hpp"
27 #include "opto/graphKit.hpp"
28 #include "opto/idealKit.hpp"
29 #include "opto/macro.hpp"
30 #include "opto/type.hpp"
31 #include "gc/g1/c2G1BSCodeGen.hpp"
32 #include "gc/g1/g1BarrierSet.hpp"
33 #include "gc/g1/g1CardTable.hpp"
34 #include "gc/g1/heapRegion.hpp"
35 #include "utilities/macros.hpp"
36
37 const TypeFunc *C2G1BSCodeGen::g1_wb_pre_Type() {
38 const Type **fields = TypeTuple::fields(2);
39 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
40 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
41 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
42
43 // create result type (range)
44 fields = TypeTuple::fields(0);
45 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
46
47 return TypeFunc::make(domain, range);
48 }
49
50 const TypeFunc *C2G1BSCodeGen::g1_wb_post_Type() {
51 const Type **fields = TypeTuple::fields(2);
52 fields[TypeFunc::Parms+0] = TypeRawPtr::NOTNULL; // Card addr
53 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
54 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
55
56 // create result type (range)
57 fields = TypeTuple::fields(0);
58 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms, fields);
59
60 return TypeFunc::make(domain, range);
61 }
62
63 #define __ ideal.
64 /*
65 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
66 * required by SATB to make sure all objects live at the start of the
67 * marking are kept alive, all reference updates need to any previous
68 * reference stored before writing.
69 *
70 * If the previous value is NULL there is no need to save the old value.
71 * References that are NULL are filtered during runtime by the barrier
72 * code to avoid unnecessary queuing.
73 *
74 * However in the case of newly allocated objects it might be possible to
75 * prove that the reference about to be overwritten is NULL during compile
76 * time and avoid adding the barrier code completely.
77 *
78 * The compiler needs to determine that the object in which a field is about
79 * to be written is newly allocated, and that no prior store to the same field
80 * has happened since the allocation.
81 *
82 * Returns true if the pre-barrier can be removed
83 */
84 bool C2G1BSCodeGen::g1_can_remove_pre_barrier(GraphKit* kit,
85 PhaseTransform* phase,
86 Node* adr,
87 BasicType bt,
88 uint adr_idx) {
89 intptr_t offset = 0;
90 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
91 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
92
93 if (offset == Type::OffsetBot) {
94 return false; // cannot unalias unless there are precise offsets
95 }
96
97 if (alloc == NULL) {
98 return false; // No allocation found
99 }
100
101 intptr_t size_in_bytes = type2aelembytes(bt);
102
103 Node* mem = kit->memory(adr_idx); // start searching here...
104
105 for (int cnt = 0; cnt < 50; cnt++) {
106
107 if (mem->is_Store()) {
108
109 Node* st_adr = mem->in(MemNode::Address);
110 intptr_t st_offset = 0;
111 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
112
113 if (st_base == NULL) {
114 break; // inscrutable pointer
115 }
116
117 // Break we have found a store with same base and offset as ours so break
118 if (st_base == base && st_offset == offset) {
119 break;
120 }
121
122 if (st_offset != offset && st_offset != Type::OffsetBot) {
123 const int MAX_STORE = BytesPerLong;
124 if (st_offset >= offset + size_in_bytes ||
125 st_offset <= offset - MAX_STORE ||
126 st_offset <= offset - mem->as_Store()->memory_size()) {
127 // Success: The offsets are provably independent.
128 // (You may ask, why not just test st_offset != offset and be done?
129 // The answer is that stores of different sizes can co-exist
130 // in the same sequence of RawMem effects. We sometimes initialize
131 // a whole 'tile' of array elements with a single jint or jlong.)
132 mem = mem->in(MemNode::Memory);
133 continue; // advance through independent store memory
134 }
135 }
136
137 if (st_base != base
138 && MemNode::detect_ptr_independence(base, alloc, st_base,
139 AllocateNode::Ideal_allocation(st_base, phase),
140 phase)) {
141 // Success: The bases are provably independent.
142 mem = mem->in(MemNode::Memory);
143 continue; // advance through independent store memory
144 }
145 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
146
147 InitializeNode* st_init = mem->in(0)->as_Initialize();
148 AllocateNode* st_alloc = st_init->allocation();
149
150 // Make sure that we are looking at the same allocation site.
151 // The alloc variable is guaranteed to not be null here from earlier check.
152 if (alloc == st_alloc) {
153 // Check that the initialization is storing NULL so that no previous store
154 // has been moved up and directly write a reference
155 Node* captured_store = st_init->find_captured_store(offset,
156 type2aelembytes(T_OBJECT),
157 phase);
158 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
159 return true;
160 }
161 }
162 }
163
164 // Unless there is an explicit 'continue', we must bail out here,
165 // because 'mem' is an inscrutable memory state (e.g., a call).
166 break;
167 }
168
169 return false;
170 }
171
172 // G1 pre/post barriers
173 void C2G1BSCodeGen::pre_barrier(GraphKit* kit,
174 bool do_load,
175 Node* ctl,
176 Node* obj,
177 Node* adr,
178 uint alias_idx,
179 Node* val,
180 const TypeOopPtr* val_type,
181 Node* pre_val,
182 BasicType bt) {
183 // Some sanity checks
184 // Note: val is unused in this routine.
185
186 if (do_load) {
187 // We need to generate the load of the previous value
188 assert(obj != NULL, "must have a base");
189 assert(adr != NULL, "where are loading from?");
190 assert(pre_val == NULL, "loaded already?");
191 assert(val_type != NULL, "need a type");
192
193 if (use_ReduceInitialCardMarks()
194 && g1_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
195 return;
196 }
197
198 } else {
199 // In this case both val_type and alias_idx are unused.
200 assert(pre_val != NULL, "must be loaded already");
201 // Nothing to be done if pre_val is null.
202 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
203 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
204 }
205 assert(bt == T_OBJECT, "or we shouldn't be here");
206
207 IdealKit ideal(kit, true);
208
209 Node* tls = __ thread(); // ThreadLocalStorage
210
211 Node* no_ctrl = NULL;
212 Node* no_base = __ top();
213 Node* zero = __ ConI(0);
214 Node* zeroX = __ ConX(0);
215
216 float likely = PROB_LIKELY(0.999);
217 float unlikely = PROB_UNLIKELY(0.999);
218
219 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
220 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
221
222 // Offsets into the thread
223 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
224 SATBMarkQueue::byte_offset_of_active());
225 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
226 SATBMarkQueue::byte_offset_of_index());
227 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
228 SATBMarkQueue::byte_offset_of_buf());
229
230 // Now the actual pointers into the thread
231 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
232 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
233 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
234
235 // Now some of the values
236 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
237
238 // if (!marking)
239 __ if_then(marking, BoolTest::ne, zero, unlikely); {
240 BasicType index_bt = TypeX_X->basic_type();
241 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
242 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
243
244 if (do_load) {
245 // load original value
246 // alias_idx correct??
247 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
248 }
249
250 // if (pre_val != NULL)
251 __ if_then(pre_val, BoolTest::ne, kit->null()); {
252 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
253
254 // is the queue for this thread full?
255 __ if_then(index, BoolTest::ne, zeroX, likely); {
256
257 // decrement the index
258 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
259
260 // Now get the buffer location we will log the previous value into and store it
261 Node *log_addr = __ AddP(no_base, buffer, next_index);
262 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
263 // update the index
264 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
265
266 } __ else_(); {
267
268 // logging buffer is full, call the runtime
269 const TypeFunc *tf = g1_wb_pre_Type();
270 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, G1BarrierSet::g1_wb_pre), "g1_wb_pre", pre_val, tls);
271 } __ end_if(); // (!index)
272 } __ end_if(); // (pre_val != NULL)
273 } __ end_if(); // (!marking)
274
275 // Final sync IdealKit and GraphKit.
276 kit->final_sync(ideal);
277 }
278
279 /*
280 * G1 similar to any GC with a Young Generation requires a way to keep track of
281 * references from Old Generation to Young Generation to make sure all live
282 * objects are found. G1 also requires to keep track of object references
283 * between different regions to enable evacuation of old regions, which is done
284 * as part of mixed collections. References are tracked in remembered sets and
285 * is continuously updated as reference are written to with the help of the
286 * post-barrier.
287 *
288 * To reduce the number of updates to the remembered set the post-barrier
289 * filters updates to fields in objects located in the Young Generation,
290 * the same region as the reference, when the NULL is being written or
291 * if the card is already marked as dirty by an earlier write.
292 *
293 * Under certain circumstances it is possible to avoid generating the
294 * post-barrier completely if it is possible during compile time to prove
295 * the object is newly allocated and that no safepoint exists between the
296 * allocation and the store.
297 *
298 * In the case of slow allocation the allocation code must handle the barrier
299 * as part of the allocation in the case the allocated object is not located
300 * in the nursery, this would happen for humongous objects. This is similar to
301 * how CMS is required to handle this case, see the comments for the method
302 * CollectedHeap::new_deferred_store_barrier and OptoRuntime::new_deferred_store_barrier.
303 * A deferred card mark is required for these objects and handled in the above
304 * mentioned methods.
305 *
306 * Returns true if the post barrier can be removed
307 */
308 bool C2G1BSCodeGen::g1_can_remove_post_barrier(GraphKit* kit,
309 PhaseTransform* phase, Node* store,
310 Node* adr) {
311 intptr_t offset = 0;
312 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
313 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
314
315 if (offset == Type::OffsetBot) {
316 return false; // cannot unalias unless there are precise offsets
317 }
318
319 if (alloc == NULL) {
320 return false; // No allocation found
321 }
322
323 // Start search from Store node
324 Node* mem = store->in(MemNode::Control);
325 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
326
327 InitializeNode* st_init = mem->in(0)->as_Initialize();
328 AllocateNode* st_alloc = st_init->allocation();
329
330 // Make sure we are looking at the same allocation
331 if (alloc == st_alloc) {
332 return true;
333 }
334 }
335
336 return false;
337 }
338
339 //
340 // Update the card table and add card address to the queue
341 //
342 void C2G1BSCodeGen::g1_mark_card(GraphKit* kit,
343 IdealKit& ideal,
344 Node* card_adr,
345 Node* oop_store,
346 uint oop_alias_idx,
347 Node* index,
348 Node* index_adr,
349 Node* buffer,
350 const TypeFunc* tf) {
351 Node* zero = __ ConI(0);
352 Node* zeroX = __ ConX(0);
353 Node* no_base = __ top();
354 BasicType card_bt = T_BYTE;
355 // Smash zero into card. MUST BE ORDERED WRT TO STORE
356 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
357
358 // Now do the queue work
359 __ if_then(index, BoolTest::ne, zeroX); {
360
361 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
362 Node* log_addr = __ AddP(no_base, buffer, next_index);
363
364 // Order, see storeCM.
365 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
366 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
367
368 } __ else_(); {
369 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, G1BarrierSet::g1_wb_post), "g1_wb_post", card_adr, __ thread());
370 } __ end_if();
371
372 }
373
374 void C2G1BSCodeGen::post_barrier(GraphKit* kit,
375 Node* ctl,
376 Node* oop_store,
377 Node* obj,
378 Node* adr,
379 uint alias_idx,
380 Node* val,
381 BasicType bt,
382 bool use_precise) {
383 // If we are writing a NULL then we need no post barrier
384
385 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
386 // Must be NULL
387 const Type* t = val->bottom_type();
388 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
389 // No post barrier if writing NULLx
390 return;
391 }
392
393 if (use_ReduceInitialCardMarks() && obj == kit->just_allocated_object(kit->control())) {
394 // We can skip marks on a freshly-allocated object in Eden.
395 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
396 // That routine informs GC to take appropriate compensating steps,
397 // upon a slow-path allocation, so as to make this card-mark
398 // elision safe.
399 return;
400 }
401
402 if (use_ReduceInitialCardMarks()
403 && g1_can_remove_post_barrier(kit, &kit->gvn(), oop_store, adr)) {
404 return;
405 }
406
407 if (!use_precise) {
408 // All card marks for a (non-array) instance are in one place:
409 adr = obj;
410 }
411 // (Else it's an array (or unknown), and we want more precise card marks.)
412 assert(adr != NULL, "");
413
414 IdealKit ideal(kit, true);
415
416 Node* tls = __ thread(); // ThreadLocalStorage
417
418 Node* no_base = __ top();
419 float unlikely = PROB_UNLIKELY(0.999);
420 Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val());
421 Node* dirty_card = __ ConI((jint)G1CardTable::dirty_card_val());
422 Node* zeroX = __ ConX(0);
423
424 const TypeFunc *tf = g1_wb_post_Type();
425
426 // Offsets into the thread
427 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
428 DirtyCardQueue::byte_offset_of_index());
429 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
430 DirtyCardQueue::byte_offset_of_buf());
431
432 // Pointers into the thread
433
434 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
435 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
436
437 // Now some values
438 // Use ctrl to avoid hoisting these values past a safepoint, which could
439 // potentially reset these fields in the JavaThread.
440 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
441 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
442
443 // Convert the store obj pointer to an int prior to doing math on it
444 // Must use ctrl to prevent "integerized oop" existing across safepoint
445 Node* cast = __ CastPX(__ ctrl(), adr);
446
447 // Divide pointer by card size
448 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
449
450 // Combine card table base and card offset
451 Node* card_adr = __ AddP(no_base, byte_map_base_node(kit), card_offset );
452
453 // If we know the value being stored does it cross regions?
454
455 if (val != NULL) {
456 // Does the store cause us to cross regions?
457
458 // Should be able to do an unsigned compare of region_size instead of
459 // and extra shift. Do we have an unsigned compare??
460 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
461 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
462
463 // if (xor_res == 0) same region so skip
464 __ if_then(xor_res, BoolTest::ne, zeroX); {
465
466 // No barrier if we are storing a NULL
467 __ if_then(val, BoolTest::ne, kit->null(), unlikely); {
468
469 // Ok must mark the card if not already dirty
470
471 // load the original value of the card
472 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
473
474 __ if_then(card_val, BoolTest::ne, young_card); {
475 kit->sync_kit(ideal);
476 kit->insert_mem_bar(Op_MemBarVolatile, oop_store);
477 __ sync_kit(kit);
478
479 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
480 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
481 g1_mark_card(kit, ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
482 } __ end_if();
483 } __ end_if();
484 } __ end_if();
485 } __ end_if();
486 } else {
487 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
488 // We don't need a barrier here if the destination is a newly allocated object
489 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
490 // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()).
491 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
492 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
493 __ if_then(card_val, BoolTest::ne, young_card); {
494 g1_mark_card(kit, ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
495 } __ end_if();
496 }
497
498 // Final sync IdealKit and GraphKit.
499 kit->final_sync(ideal);
500 }
501
502 // Helper that guards and inserts a pre-barrier.
503 void C2G1BSCodeGen::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
504 Node* pre_val, bool need_mem_bar) {
505 // We could be accessing the referent field of a reference object. If so, when G1
506 // is enabled, we need to log the value in the referent field in an SATB buffer.
507 // This routine performs some compile time filters and generates suitable
508 // runtime filters that guard the pre-barrier code.
509 // Also add memory barrier for non volatile load from the referent field
510 // to prevent commoning of loads across safepoint.
511
512 // Some compile time checks.
513
514 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
515 const TypeX* otype = offset->find_intptr_t_type();
516 if (otype != NULL && otype->is_con() &&
517 otype->get_con() != java_lang_ref_Reference::referent_offset) {
518 // Constant offset but not the reference_offset so just return
519 return;
520 }
521
522 // We only need to generate the runtime guards for instances.
523 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
524 if (btype != NULL) {
525 if (btype->isa_aryptr()) {
526 // Array type so nothing to do
527 return;
528 }
529
530 const TypeInstPtr* itype = btype->isa_instptr();
531 if (itype != NULL) {
532 // Can the klass of base_oop be statically determined to be
533 // _not_ a sub-class of Reference and _not_ Object?
534 ciKlass* klass = itype->klass();
535 if ( klass->is_loaded() &&
536 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
537 !kit->env()->Object_klass()->is_subtype_of(klass)) {
538 return;
539 }
540 }
541 }
542
543 // The compile time filters did not reject base_oop/offset so
544 // we need to generate the following runtime filters
545 //
546 // if (offset == java_lang_ref_Reference::_reference_offset) {
547 // if (instance_of(base, java.lang.ref.Reference)) {
548 // pre_barrier(_, pre_val, ...);
549 // }
550 // }
551
552 float likely = PROB_LIKELY( 0.999);
553 float unlikely = PROB_UNLIKELY(0.999);
554
555 IdealKit ideal(kit);
556
557 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
558
559 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
560 // Update graphKit memory and control from IdealKit.
561 kit->sync_kit(ideal);
562
563 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
564 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
565
566 // Update IdealKit memory and control from graphKit.
567 __ sync_kit(kit);
568
569 Node* one = __ ConI(1);
570 // is_instof == 0 if base_oop == NULL
571 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
572
573 // Update graphKit from IdeakKit.
574 kit->sync_kit(ideal);
575
576 // Use the pre-barrier to record the value in the referent field
577 pre_barrier(kit, false /* do_load */,
578 __ ctrl(),
579 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
580 pre_val /* pre_val */,
581 T_OBJECT);
582 if (need_mem_bar) {
583 // Add memory barrier to prevent commoning reads from this field
584 // across safepoint since GC can change its value.
585 kit->insert_mem_bar(Op_MemBarCPUOrder);
586 }
587 // Update IdealKit from graphKit.
588 __ sync_kit(kit);
589
590 } __ end_if(); // _ref_type != ref_none
591 } __ end_if(); // offset == referent_offset
592
593 // Final sync IdealKit and GraphKit.
594 kit->final_sync(ideal);
595 }
596
597 #undef __
598
599 Node* C2G1BSCodeGen::load_at_resolved(GraphKit* kit, Node* obj, Node* adr, const TypePtr* adr_type, const Type* val_type, BasicType bt, C2DecoratorSet decorators) {
600 bool is_obj = bt == T_OBJECT || bt == T_ARRAY;
601 bool mismatched = (decorators & C2_MISMATCHED) != 0;
602 bool anonymous = (decorators & C2_ACCESS_ON_ANONYMOUS) != 0;
603 bool on_heap = (decorators & C2_ACCESS_ON_HEAP) != 0;
604 bool on_weak = (decorators & C2_ACCESS_ON_WEAK) != 0;
605 bool is_relaxed = (decorators & C2_MO_RELAXED) != 0;
606 bool need_cpu_mem_bar = !is_relaxed || mismatched || !on_heap;
607
608 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : kit->top();
609 Node* load = C2CardTableModRefBSCodeGen::load_at_resolved(kit, obj, adr, adr_type, val_type, bt, decorators);
610
611 // If we are reading the value of the referent field of a Reference
612 // object (either by using Unsafe directly or through reflection)
613 // then, if G1 is enabled, we need to record the referent in an
614 // SATB log buffer using the pre-barrier mechanism.
615 // Also we need to add memory barrier to prevent commoning reads
616 // from this field across safepoint since GC can change its value.
617 bool need_read_barrier = on_heap && (on_weak || (anonymous && offset != kit->top() && obj != kit->top()));
618
619 if (!is_obj || !need_read_barrier) {
620 return load;
621 }
622
623 if (on_weak) {
624 // Use the pre-barrier to record the value in the referent field
625 pre_barrier(kit, false /* do_load */,
626 kit->control(),
627 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
628 load /* pre_val */,
629 T_OBJECT);
630 // Add memory barrier to prevent commoning reads from this field
631 // across safepoint since GC can change its value.
632 kit->insert_mem_bar(Op_MemBarCPUOrder);
633 } else if (anonymous) {
634 // We do not require a mem bar inside pre_barrier if need_mem_bar
635 // is set: the barriers would be emitted by us.
636 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
637 }
638
639 return load;
640 }
641
642 bool C2G1BSCodeGen::is_gc_barrier_node(Node* node) {
643 if (C2CardTableModRefBSCodeGen::is_gc_barrier_node(node)) {
644 return true;
645 }
646 if (node->Opcode() != Op_CallLeaf) {
647 return false;
648 }
649 CallLeafNode *call = node->as_CallLeaf();
650 if (call->_name == NULL) {
651 return false;
652 }
653
654 return strcmp(call->_name, "g1_wb_pre") == 0 || strcmp(call->_name, "g1_wb_post") == 0;
655 }
656
657 void C2G1BSCodeGen::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) {
658 assert(node->Opcode() == Op_CastP2X, "ConvP2XNode required");
659 assert(node->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
660 // It could be only one user, URShift node, in Object.clone() intrinsic
661 // but the new allocation is passed to arraycopy stub and it could not
662 // be scalar replaced. So we don't check the case.
663
664 // An other case of only one user (Xor) is when the value check for NULL
665 // in G1 post barrier is folded after CCP so the code which used URShift
666 // is removed.
667
668 // Take Region node before eliminating post barrier since it also
669 // eliminates CastP2X node when it has only one user.
670 Node* this_region = node->in(0);
671 assert(this_region != NULL, "");
672
673 // Remove G1 post barrier.
674
675 // Search for CastP2X->Xor->URShift->Cmp path which
676 // checks if the store done to a different from the value's region.
677 // And replace Cmp with #0 (false) to collapse G1 post barrier.
678 Node* xorx = node->find_out_with(Op_XorX);
679 if (xorx != NULL) {
680 Node* shift = xorx->unique_out();
681 Node* cmpx = shift->unique_out();
682 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
683 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
684 "missing region check in G1 post barrier");
685 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
686
687 // Remove G1 pre barrier.
688
689 // Search "if (marking != 0)" check and set it to "false".
690 // There is no G1 pre barrier if previous stored value is NULL
691 // (for example, after initialization).
692 if (this_region->is_Region() && this_region->req() == 3) {
693 int ind = 1;
694 if (!this_region->in(ind)->is_IfFalse()) {
695 ind = 2;
696 }
697 if (this_region->in(ind)->is_IfFalse()) {
698 Node* bol = this_region->in(ind)->in(0)->in(1);
699 assert(bol->is_Bool(), "");
700 cmpx = bol->in(1);
701 if (bol->as_Bool()->_test._test == BoolTest::ne &&
702 cmpx->is_Cmp() && cmpx->in(2) == macro->intcon(0) &&
703 cmpx->in(1)->is_Load()) {
704 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
705 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
706 SATBMarkQueue::byte_offset_of_active());
707 if (adr->is_AddP() && adr->in(AddPNode::Base) == macro->top() &&
708 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
709 adr->in(AddPNode::Offset) == macro->MakeConX(marking_offset)) {
710 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
711 }
712 }
713 }
714 }
715 } else {
716 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
717 // This is a G1 post barrier emitted by the Object.clone() intrinsic.
718 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card
719 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier.
720 Node* shift = node->find_out_with(Op_URShiftX);
721 assert(shift != NULL, "missing G1 post barrier");
722 Node* addp = shift->unique_out();
723 Node* load = addp->find_out_with(Op_LoadB);
724 assert(load != NULL, "missing G1 post barrier");
725 Node* cmpx = load->unique_out();
726 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
727 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
728 "missing card value check in G1 post barrier");
729 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
730 // There is no G1 pre barrier in this case
731 }
732 // Now CastP2X can be removed since it is used only on dead path
733 // which currently still alive until igvn optimize it.
734 assert(node->outcnt() == 0 || node->unique_out()->Opcode() == Op_URShiftX, "");
735 macro->replace_node(node, macro->top());
736 }