1 /*
2 * Copyright (c) 2018, 2020, 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 #define __ ideal.
69 /*
70 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
71 * required by SATB to make sure all objects live at the start of the
72 * marking are kept alive, all reference updates need to any previous
73 * reference stored before writing.
74 *
75 * If the previous value is NULL there is no need to save the old value.
76 * References that are NULL are filtered during runtime by the barrier
77 * code to avoid unnecessary queuing.
78 *
79 * However in the case of newly allocated objects it might be possible to
80 * prove that the reference about to be overwritten is NULL during compile
81 * time and avoid adding the barrier code completely.
82 *
83 * The compiler needs to determine that the object in which a field is about
84 * to be written is newly allocated, and that no prior store to the same field
85 * has happened since the allocation.
86 *
87 * Returns true if the pre-barrier can be removed
88 */
89 bool G1BarrierSetC2::g1_can_remove_pre_barrier(GraphKit* kit,
90 PhaseTransform* phase,
91 Node* adr,
92 BasicType bt,
93 uint adr_idx) const {
94 intptr_t offset = 0;
95 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
96 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
97
98 if (offset == Type::OffsetBot) {
99 return false; // cannot unalias unless there are precise offsets
100 }
101
102 if (alloc == NULL) {
103 return false; // No allocation found
104 }
105
106 intptr_t size_in_bytes = type2aelembytes(bt);
107
108 Node* mem = kit->memory(adr_idx); // start searching here...
109
110 for (int cnt = 0; cnt < 50; cnt++) {
111
112 if (mem->is_Store()) {
113
114 Node* st_adr = mem->in(MemNode::Address);
115 intptr_t st_offset = 0;
116 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
117
118 if (st_base == NULL) {
119 break; // inscrutable pointer
120 }
121
122 // Break we have found a store with same base and offset as ours so break
123 if (st_base == base && st_offset == offset) {
124 break;
125 }
126
127 if (st_offset != offset && st_offset != Type::OffsetBot) {
128 const int MAX_STORE = BytesPerLong;
129 if (st_offset >= offset + size_in_bytes ||
130 st_offset <= offset - MAX_STORE ||
131 st_offset <= offset - mem->as_Store()->memory_size()) {
132 // Success: The offsets are provably independent.
133 // (You may ask, why not just test st_offset != offset and be done?
134 // The answer is that stores of different sizes can co-exist
135 // in the same sequence of RawMem effects. We sometimes initialize
136 // a whole 'tile' of array elements with a single jint or jlong.)
137 mem = mem->in(MemNode::Memory);
138 continue; // advance through independent store memory
139 }
140 }
141
142 if (st_base != base
143 && MemNode::detect_ptr_independence(base, alloc, st_base,
144 AllocateNode::Ideal_allocation(st_base, phase),
145 phase)) {
146 // Success: The bases are provably independent.
147 mem = mem->in(MemNode::Memory);
148 continue; // advance through independent store memory
149 }
150 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
151
152 InitializeNode* st_init = mem->in(0)->as_Initialize();
153 AllocateNode* st_alloc = st_init->allocation();
154
155 // Make sure that we are looking at the same allocation site.
156 // The alloc variable is guaranteed to not be null here from earlier check.
157 if (alloc == st_alloc) {
158 // Check that the initialization is storing NULL so that no previous store
159 // has been moved up and directly write a reference
160 Node* captured_store = st_init->find_captured_store(offset,
161 type2aelembytes(T_OBJECT),
162 phase);
163 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
164 return true;
165 }
166 }
167 }
168
169 // Unless there is an explicit 'continue', we must bail out here,
170 // because 'mem' is an inscrutable memory state (e.g., a call).
171 break;
172 }
173
174 return false;
175 }
176
177 // G1 pre/post barriers
178 void G1BarrierSetC2::pre_barrier(GraphKit* kit,
179 bool do_load,
180 Node* ctl,
181 Node* obj,
182 Node* adr,
183 uint alias_idx,
184 Node* val,
185 const TypeOopPtr* val_type,
186 Node* pre_val,
187 BasicType bt) const {
188 // Some sanity checks
189 // Note: val is unused in this routine.
190
191 if (do_load) {
192 // We need to generate the load of the previous value
193 assert(obj != NULL, "must have a base");
194 assert(adr != NULL, "where are loading from?");
195 assert(pre_val == NULL, "loaded already?");
196 assert(val_type != NULL, "need a type");
197
198 if (use_ReduceInitialCardMarks()
199 && g1_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
200 return;
201 }
202
203 } else {
204 // In this case both val_type and alias_idx are unused.
205 assert(pre_val != NULL, "must be loaded already");
206 // Nothing to be done if pre_val is null.
207 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
208 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
209 }
210 assert(bt == T_OBJECT, "or we shouldn't be here");
211
212 IdealKit ideal(kit, true);
213
214 Node* tls = __ thread(); // ThreadLocalStorage
215
216 Node* no_base = __ top();
217 Node* zero = __ ConI(0);
218 Node* zeroX = __ ConX(0);
219
220 float likely = PROB_LIKELY(0.999);
221 float unlikely = PROB_UNLIKELY(0.999);
222
223 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
224 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
225
226 // Offsets into the thread
227 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
228 const int index_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset());
229 const int buffer_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset());
230
231 // Now the actual pointers into the thread
232 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
233 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
234 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
235
236 // Now some of the values
237 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
238
239 // if (!marking)
240 __ if_then(marking, BoolTest::ne, zero, unlikely); {
241 BasicType index_bt = TypeX_X->basic_type();
242 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
243 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
244
245 if (do_load) {
246 // load original value
247 // alias_idx correct??
248 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
249 }
250
251 // if (pre_val != NULL)
252 __ if_then(pre_val, BoolTest::ne, kit->null()); {
253 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
254
255 // is the queue for this thread full?
256 __ if_then(index, BoolTest::ne, zeroX, likely); {
257
258 // decrement the index
259 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
260
261 // Now get the buffer location we will log the previous value into and store it
262 Node *log_addr = __ AddP(no_base, buffer, next_index);
263 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
264 // update the index
265 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
266
267 } __ else_(); {
268
269 // logging buffer is full, call the runtime
270 const TypeFunc *tf = write_ref_field_pre_entry_Type();
271 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), "write_ref_field_pre_entry", pre_val, tls);
272 } __ end_if(); // (!index)
273 } __ end_if(); // (pre_val != NULL)
274 } __ end_if(); // (!marking)
275
276 // Final sync IdealKit and GraphKit.
277 kit->final_sync(ideal);
278 }
279
280 /*
281 * G1 similar to any GC with a Young Generation requires a way to keep track of
282 * references from Old Generation to Young Generation to make sure all live
283 * objects are found. G1 also requires to keep track of object references
284 * between different regions to enable evacuation of old regions, which is done
285 * as part of mixed collections. References are tracked in remembered sets and
286 * is continuously updated as reference are written to with the help of the
287 * post-barrier.
288 *
289 * To reduce the number of updates to the remembered set the post-barrier
290 * filters updates to fields in objects located in the Young Generation,
291 * the same region as the reference, when the NULL is being written or
292 * if the card is already marked as dirty by an earlier write.
293 *
294 * Under certain circumstances it is possible to avoid generating the
295 * post-barrier completely if it is possible during compile time to prove
296 * the object is newly allocated and that no safepoint exists between the
297 * allocation and the store.
298 *
299 * In the case of slow allocation the allocation code must handle the barrier
300 * as part of the allocation in the case the allocated object is not located
301 * in the nursery; this would happen for humongous objects.
302 *
303 * Returns true if the post barrier can be removed
304 */
305 bool G1BarrierSetC2::g1_can_remove_post_barrier(GraphKit* kit,
306 PhaseTransform* phase, Node* store,
307 Node* adr) const {
308 intptr_t offset = 0;
309 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
310 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
311
312 if (offset == Type::OffsetBot) {
313 return false; // cannot unalias unless there are precise offsets
314 }
315
316 if (alloc == NULL) {
317 return false; // No allocation found
318 }
319
320 // Start search from Store node
321 Node* mem = store->in(MemNode::Control);
322 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
323
324 InitializeNode* st_init = mem->in(0)->as_Initialize();
325 AllocateNode* st_alloc = st_init->allocation();
326
327 // Make sure we are looking at the same allocation
328 if (alloc == st_alloc) {
329 return true;
330 }
331 }
332
333 return false;
334 }
335
336 //
337 // Update the card table and add card address to the queue
338 //
339 void G1BarrierSetC2::g1_mark_card(GraphKit* kit,
340 IdealKit& ideal,
341 Node* card_adr,
342 Node* oop_store,
343 uint oop_alias_idx,
344 Node* index,
345 Node* index_adr,
346 Node* buffer,
347 const TypeFunc* tf) const {
348 Node* zero = __ ConI(0);
349 Node* zeroX = __ ConX(0);
350 Node* no_base = __ top();
351 BasicType card_bt = T_BYTE;
352 // Smash zero into card. MUST BE ORDERED WRT TO STORE
353 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
354
355 // Now do the queue work
356 __ if_then(index, BoolTest::ne, zeroX); {
357
358 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
359 Node* log_addr = __ AddP(no_base, buffer, next_index);
360
361 // Order, see storeCM.
362 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
363 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
364
365 } __ else_(); {
366 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_post_entry), "write_ref_field_post_entry", card_adr, __ thread());
367 } __ end_if();
368
369 }
370
371 void G1BarrierSetC2::post_barrier(GraphKit* kit,
372 Node* ctl,
373 Node* oop_store,
374 Node* obj,
375 Node* adr,
376 uint alias_idx,
377 Node* val,
378 BasicType bt,
379 bool use_precise) const {
380 // If we are writing a NULL then we need no post barrier
381
382 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
383 // Must be NULL
384 const Type* t = val->bottom_type();
385 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
386 // No post barrier if writing NULLx
387 return;
388 }
389
390 if (use_ReduceInitialCardMarks() && obj == kit->just_allocated_object(kit->control())) {
391 // We can skip marks on a freshly-allocated object in Eden.
392 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
393 // That routine informs GC to take appropriate compensating steps,
394 // upon a slow-path allocation, so as to make this card-mark
395 // elision safe.
396 return;
397 }
398
399 if (use_ReduceInitialCardMarks()
400 && g1_can_remove_post_barrier(kit, &kit->gvn(), oop_store, adr)) {
401 return;
402 }
403
404 if (!use_precise) {
405 // All card marks for a (non-array) instance are in one place:
406 adr = obj;
407 }
408 // (Else it's an array (or unknown), and we want more precise card marks.)
409 assert(adr != NULL, "");
410
411 IdealKit ideal(kit, true);
412
413 Node* tls = __ thread(); // ThreadLocalStorage
414
415 BarrierSet* bs = BarrierSet::barrier_set();
416 CardTableBarrierSet* ctbs = barrier_set_cast<CardTableBarrierSet>(bs);
417 CardTable* ct = ctbs->card_table();
418
419 Node* no_base = __ top();
420 float likely = PROB_LIKELY_MAG(3);
421 float unlikely = PROB_UNLIKELY_MAG(3);
422 Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val());
423 Node* dirty_card = __ ConI((jint)G1CardTable::dirty_card_val());
424 Node* zeroX = __ ConX(0);
425
426 const TypeFunc *tf = write_ref_field_post_entry_Type();
427
428 // Offsets into the thread
429 const int index_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset());
430 const int buffer_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset());
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, likely); {
465
466 // if ((unsigned)(card_offset - low_map_offset) >= (high_map_offset - low_map_offset)) stack allocated object, so skip
467 if (kit->C->do_stack_allocation()) {
468 state()->add_enqueue_barrier(static_cast<CastP2XNode*>(cast));
469 Node* low_off = kit->longcon(ct->byte_map_bottom_offset());
470 Node* delta_off = kit->longcon(ct->byte_map_top_offset() - ct->byte_map_bottom_offset());
471 Node* sub_off = __ SubL(cast, low_off);
472
473 __ uif_then(sub_off, BoolTest::le, delta_off, likely); } {
474
475 // No barrier if we are storing a NULL
476 __ if_then(val, BoolTest::ne, kit->null(), likely); {
477
478 // Ok must mark the card if not already dirty
479
480 // load the original value of the card
481 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
482
483 __ if_then(card_val, BoolTest::ne, young_card, unlikely); {
484 kit->sync_kit(ideal);
485 kit->insert_mem_bar(Op_MemBarVolatile, oop_store);
486 __ sync_kit(kit);
487
488 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
489 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
490 g1_mark_card(kit, ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
491 } __ end_if();
492 } __ end_if();
493 } __ end_if();
494 } if (kit->C->do_stack_allocation()) {
495 __ end_if();
496 }
497 } __ end_if();
498 } else {
499 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
500 // We don't need a barrier here if the destination is a newly allocated object
501 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
502 // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()).
503 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
504
505 // if ((unsigned)(card_offset - low_map_offset) >= (high_map_offset - low_map_offset)) stack allocated object, so skip
506 if (kit->C->do_stack_allocation()) {
507 state()->add_enqueue_barrier(static_cast<CastP2XNode*>(cast));
508 Node* low_off = kit->longcon(ct->byte_map_bottom_offset());
509 Node* delta_off = kit->longcon(ct->byte_map_top_offset() - ct->byte_map_bottom_offset());
510 Node* sub_off = __ SubL(cast, low_off);
511
512 __ uif_then(sub_off, BoolTest::le, delta_off, likely); } {
513
514 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
515 __ if_then(card_val, BoolTest::ne, young_card); {
516 g1_mark_card(kit, ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
517 } __ end_if();
518
519 } if (kit->C->do_stack_allocation()) {
520 __ end_if();
521 }
522 }
523
524 // Final sync IdealKit and GraphKit.
525 kit->final_sync(ideal);
526 }
527
528 // Helper that guards and inserts a pre-barrier.
529 void G1BarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
530 Node* pre_val, bool need_mem_bar) const {
531 // We could be accessing the referent field of a reference object. If so, when G1
532 // is enabled, we need to log the value in the referent field in an SATB buffer.
533 // This routine performs some compile time filters and generates suitable
534 // runtime filters that guard the pre-barrier code.
535 // Also add memory barrier for non volatile load from the referent field
536 // to prevent commoning of loads across safepoint.
537
538 // Some compile time checks.
539
540 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
541 const TypeX* otype = offset->find_intptr_t_type();
542 if (otype != NULL && otype->is_con() &&
543 otype->get_con() != java_lang_ref_Reference::referent_offset()) {
544 // Constant offset but not the reference_offset so just return
545 return;
546 }
547
548 // We only need to generate the runtime guards for instances.
549 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
550 if (btype != NULL) {
551 if (btype->isa_aryptr()) {
552 // Array type so nothing to do
553 return;
554 }
555
556 const TypeInstPtr* itype = btype->isa_instptr();
557 if (itype != NULL) {
558 // Can the klass of base_oop be statically determined to be
559 // _not_ a sub-class of Reference and _not_ Object?
560 ciKlass* klass = itype->klass();
561 if ( klass->is_loaded() &&
562 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
563 !kit->env()->Object_klass()->is_subtype_of(klass)) {
564 return;
565 }
566 }
567 }
568
569 // The compile time filters did not reject base_oop/offset so
570 // we need to generate the following runtime filters
571 //
572 // if (offset == java_lang_ref_Reference::_reference_offset) {
573 // if (instance_of(base, java.lang.ref.Reference)) {
574 // pre_barrier(_, pre_val, ...);
575 // }
576 // }
577
578 float likely = PROB_LIKELY( 0.999);
579 float unlikely = PROB_UNLIKELY(0.999);
580
581 IdealKit ideal(kit);
582
583 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset());
584
585 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
586 // Update graphKit memory and control from IdealKit.
587 kit->sync_kit(ideal);
588
589 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
590 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
591
592 // Update IdealKit memory and control from graphKit.
593 __ sync_kit(kit);
594
595 Node* one = __ ConI(1);
596 // is_instof == 0 if base_oop == NULL
597 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
598
599 // Update graphKit from IdeakKit.
600 kit->sync_kit(ideal);
601
602 // Use the pre-barrier to record the value in the referent field
603 pre_barrier(kit, false /* do_load */,
604 __ ctrl(),
605 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
606 pre_val /* pre_val */,
607 T_OBJECT);
608 if (need_mem_bar) {
609 // Add memory barrier to prevent commoning reads from this field
610 // across safepoint since GC can change its value.
611 kit->insert_mem_bar(Op_MemBarCPUOrder);
612 }
613 // Update IdealKit from graphKit.
614 __ sync_kit(kit);
615
616 } __ end_if(); // _ref_type != ref_none
617 } __ end_if(); // offset == referent_offset
618
619 // Final sync IdealKit and GraphKit.
620 kit->final_sync(ideal);
621 }
622
623 #undef __
624
625 Node* G1BarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
626 DecoratorSet decorators = access.decorators();
627 Node* adr = access.addr().node();
628 Node* obj = access.base();
629
630 bool anonymous = (decorators & C2_UNSAFE_ACCESS) != 0;
631 bool mismatched = (decorators & C2_MISMATCHED) != 0;
632 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
633 bool in_heap = (decorators & IN_HEAP) != 0;
634 bool in_native = (decorators & IN_NATIVE) != 0;
635 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
636 bool is_unordered = (decorators & MO_UNORDERED) != 0;
637 bool is_mixed = !in_heap && !in_native;
638 bool need_cpu_mem_bar = !is_unordered || mismatched || is_mixed;
639
640 Node* top = Compile::current()->top();
641 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
642 Node* load = CardTableBarrierSetC2::load_at_resolved(access, val_type);
643
644 // If we are reading the value of the referent field of a Reference
645 // object (either by using Unsafe directly or through reflection)
646 // then, if G1 is enabled, we need to record the referent in an
647 // SATB log buffer using the pre-barrier mechanism.
648 // Also we need to add memory barrier to prevent commoning reads
649 // from this field across safepoint since GC can change its value.
650 bool need_read_barrier = in_heap && (on_weak ||
651 (unknown && offset != top && obj != top));
652
653 if (!access.is_oop() || !need_read_barrier) {
654 return load;
655 }
656
657 assert(access.is_parse_access(), "entry not supported at optimization time");
658 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
659 GraphKit* kit = parse_access.kit();
660
661 if (on_weak) {
662 // Use the pre-barrier to record the value in the referent field
663 pre_barrier(kit, false /* do_load */,
664 kit->control(),
665 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
666 load /* pre_val */, T_OBJECT);
667 // Add memory barrier to prevent commoning reads from this field
668 // across safepoint since GC can change its value.
669 kit->insert_mem_bar(Op_MemBarCPUOrder);
670 } else if (unknown) {
671 // We do not require a mem bar inside pre_barrier if need_mem_bar
672 // is set: the barriers would be emitted by us.
673 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
674 }
675
676 return load;
677 }
678
679 bool G1BarrierSetC2::is_gc_barrier_node(Node* node) const {
680 if (CardTableBarrierSetC2::is_gc_barrier_node(node)) {
681 return true;
682 }
683 if (node->Opcode() != Op_CallLeaf) {
684 return false;
685 }
686 CallLeafNode *call = node->as_CallLeaf();
687 if (call->_name == NULL) {
688 return false;
689 }
690
691 return strcmp(call->_name, "write_ref_field_pre_entry") == 0 || strcmp(call->_name, "write_ref_field_post_entry") == 0;
692 }
693
694 bool G1BarrierSetC2::process_barrier_node(Node* node, PhaseIterGVN& igvn) const {
695 assert(node->Opcode() == Op_CastP2X, "ConvP2XNode required");
696
697 // Must have a control node
698 if (node->in(0) == NULL) {
699 return false;
700 }
701
702 // Search for CastP2X->Xor->URShift->Cmp path which
703 // checks if the store done to a different from the value's region.
704 Node* xorx = node->find_out_with(Op_XorX);
705 BoolNode* bool_node = NULL;
706
707 if (xorx != NULL) {
708
709 Node* shift = shift = xorx->unique_out();
710 Node* cmpx = shift->unique_out();
711
712 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
713 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
714 "missing region check in G1 post barrier");
715
716 Node* bol = cmpx->unique_out();
717 assert(bol->unique_out()->is_If(), "should find if after the bool node");
718 Node* if_node = bol->unique_out();
719 Node* if_true = if_node->find_out_with(Op_IfTrue);
720 assert(if_true != NULL, "there should be false projection");
721
722 Node* iff_check = if_true->find_out_with(Op_If);
723 // Not a barrier with bound check
724 if (iff_check == NULL) {
725 return false;
726 }
727
728 Node* iff_check_in_1_node = iff_check->in(1);
729 if (!iff_check_in_1_node->is_Bool()) {
730 return false;
731 }
732 bool_node = iff_check_in_1_node->as_Bool();
733
734 } else {
735 // this "could" be the the path followed when !use_ReduceInitialCardMarks() is
736 // used or when the two sides of the barrier are scalar replaced
737 //assert(false, "we managed to get here!!! process_barrier_node");
738 Node *addl_node = node->find_out_with(Op_AddL);
739 if (addl_node == NULL) {
740 return false;
741 }
742
743 Node* cmpx = addl_node->unique_out();
744 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
745 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::le,
746 "missing region check in G1 post barrier");
747
748 bool_node = cmpx->unique_out()->as_Bool();
749 }
750
751 if (bool_node->_test._test != BoolTest::le) {
752 return false;
753 }
754
755 // the input to the bool is the CMPX
756 Node* bool_node_in_1_node = bool_node->in(1);
757 if (!bool_node_in_1_node->is_Cmp()) {
758 return false;
759 }
760 CmpNode* cmp_node = bool_node_in_1_node->as_Cmp();
761
762 // the input to the CMPX is the card_table_top_offset constant
763 Node* cmp_node_in_2_node = cmp_node->in(2);
764 if (!cmp_node_in_2_node->is_Con()) {
765 return false;
766 }
767
768 BarrierSet* bs = BarrierSet::barrier_set();
769 CardTableBarrierSet* ctbs = barrier_set_cast<CardTableBarrierSet>(bs);
770 CardTable* ct = ctbs->card_table();
771 size_t constant = ct->byte_map_top_offset() - ct->byte_map_bottom_offset();
772
773 // Check that the input to this CMP node is the expected constant
774 const TypeX* otype = cmp_node_in_2_node->find_intptr_t_type();
775 if (otype != NULL && otype->is_con() &&
776 size_t(otype->get_con()) != constant) {
777 // Constant offset but not the card table size constant so just return
778 return false;
779 }
780
781 // we can't change the compare or the constant so create a new constant(0) and replace the variable
782 Node* cmp_node_in_1_node = cmp_node->in(1);
783 ConNode* zeroConstant_node = igvn.makecon(TypeX_ZERO);
784 if (cmp_node_in_1_node->_idx == zeroConstant_node->_idx) {
785 // we can get here via different nodes - but we only want to change the input once
786 return false;
787 }
788
789 igvn.rehash_node_delayed(cmp_node);
790 int numReplaced = cmp_node->replace_edge(cmp_node_in_1_node, zeroConstant_node);
791 assert(numReplaced == 1, "Failed to replace the card_offset with Conx(0)");
792 return true;
793 }
794
795 void G1BarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const {
796 assert(node->Opcode() == Op_CastP2X, "ConvP2XNode required");
797 assert(node->outcnt() <= 3, "expects 1, 2 or 3 users: Xor, URShift and SubL nodes");
798 // It could be only one user, URShift node, in Object.clone() intrinsic
799 // but the new allocation is passed to arraycopy stub and it could not
800 // be scalar replaced. So we don't check the case.
801
802 // Certain loop optimisations may introduce a CastP2X node with
803 // ConvL2I in case of an AllocateArray op. Check for that case
804 // here and do not attempt to eliminate it as write barrier.
805 if (macro->C->do_stack_allocation() && !state()->is_a_barrier(static_cast<CastP2XNode*>(node))) {
806 return;
807 }
808
809 // An other case of only one user (Xor) is when the value check for NULL
810 // in G1 post barrier is folded after CCP so the code which used URShift
811 // is removed.
812
813 // Take Region node before eliminating post barrier since it also
814 // eliminates CastP2X node when it has only one user.
815 Node* this_region = node->in(0);
816 assert(this_region != NULL, "");
817
818 // Remove G1 post barrier.
819
820 // Search for CastP2X->Xor->URShift->Cmp path which
821 // checks if the store done to a different from the value's region.
822 // And replace Cmp with #0 (false) to collapse G1 post barrier.
823 Node* xorx = node->find_out_with(Op_XorX);
824 if (xorx != NULL) {
825 Node* shift = xorx->unique_out();
826 Node* cmpx = shift->unique_out();
827 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
828 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
829 "missing region check in G1 post barrier");
830 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
831
832 // Remove G1 pre barrier.
833
834 // Search "if (marking != 0)" check and set it to "false".
835 // There is no G1 pre barrier if previous stored value is NULL
836 // (for example, after initialization).
837 if (this_region->is_Region() && this_region->req() == 3) {
838 int ind = 1;
839 if (!this_region->in(ind)->is_IfFalse()) {
840 ind = 2;
841 }
842 if (this_region->in(ind)->is_IfFalse() &&
843 this_region->in(ind)->in(0)->Opcode() == Op_If) {
844 Node* bol = this_region->in(ind)->in(0)->in(1);
845 assert(bol->is_Bool(), "");
846 cmpx = bol->in(1);
847 if (bol->as_Bool()->_test._test == BoolTest::ne &&
848 cmpx->is_Cmp() && cmpx->in(2) == macro->intcon(0) &&
849 cmpx->in(1)->is_Load()) {
850 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
851 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
852 if (adr->is_AddP() && adr->in(AddPNode::Base) == macro->top() &&
853 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
854 adr->in(AddPNode::Offset) == macro->MakeConX(marking_offset)) {
855 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
856 }
857 }
858 }
859 }
860 } else {
861 // In a scenario where the two sides of the barrier are scalar replaced
862 // or stack allocated, the XorX node will be visited more than once, because
863 // both edges will be CastP2X nodes from two distinct allocates. In certain
864 // instances, the removal of the CastP2X node will result in removal of the
865 // XorX node, causing the assert below to be hit when eliminate_gc_barrier is
866 // called for the second node.
867 // assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
868
869 // This is a G1 post barrier emitted by the Object.clone() intrinsic.
870 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card
871 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier.
872 Node* shift = node->find_out_with(Op_URShiftX);
873 assert(shift != NULL, "missing G1 post barrier");
874 Node* addp = shift->unique_out();
875 Node* load = addp->find_out_with(Op_LoadB);
876 assert(load != NULL, "missing G1 post barrier");
877 Node* cmpx = load->unique_out();
878 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
879 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
880 "missing card value check in G1 post barrier");
881 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
882 // There is no G1 pre barrier in this case
883 }
884 // Now CastP2X can be removed since it is used only on dead path
885 // which currently still alive until igvn optimize it.
886 // TODO: fix this following assert becuase of SUBL
887 // assert(node->outcnt() == 0 || node->unique_out()->Opcode() == Op_URShiftX, "");
888 macro->replace_node(node, macro->top());
889
890 // Remove this node from our state
891 state()->remove_enqueue_barrier(static_cast<CastP2XNode*>(node));
892 }
893
894 Node* G1BarrierSetC2::step_over_gc_barrier(Node* c) const {
895 if (!use_ReduceInitialCardMarks() &&
896 c != NULL && c->is_Region() && c->req() == 3) {
897 for (uint i = 1; i < c->req(); i++) {
898 if (c->in(i) != NULL && c->in(i)->is_Region() &&
899 c->in(i)->req() == 3) {
900 Node* r = c->in(i);
901 for (uint j = 1; j < r->req(); j++) {
902 if (r->in(j) != NULL && r->in(j)->is_Proj() &&
903 r->in(j)->in(0) != NULL &&
904 r->in(j)->in(0)->Opcode() == Op_CallLeaf &&
905 r->in(j)->in(0)->as_Call()->entry_point() == CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_post_entry)) {
906 Node* call = r->in(j)->in(0);
907 c = c->in(i == 1 ? 2 : 1);
908 if (c != NULL) {
909 c = c->in(0);
910 if (c != NULL) {
911 c = c->in(0);
912 assert(call->in(0) == NULL ||
913 call->in(0)->in(0) == NULL ||
914 call->in(0)->in(0)->in(0) == NULL ||
915 call->in(0)->in(0)->in(0)->in(0) == NULL ||
916 call->in(0)->in(0)->in(0)->in(0)->in(0) == NULL ||
917 c == call->in(0)->in(0)->in(0)->in(0)->in(0), "bad barrier shape");
918 return c;
919 }
920 }
921 }
922 }
923 }
924 }
925 }
926 return c;
927 }
928
929 #ifdef ASSERT
930 void G1BarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
931 if (phase != BarrierSetC2::BeforeCodeGen) {
932 return;
933 }
934 // Verify G1 pre-barriers
935 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
936
937 ResourceArea *area = Thread::current()->resource_area();
938 Unique_Node_List visited(area);
939 Node_List worklist(area);
940 // We're going to walk control flow backwards starting from the Root
941 worklist.push(compile->root());
942 while (worklist.size() > 0) {
943 Node* x = worklist.pop();
944 if (x == NULL || x == compile->top()) continue;
945 if (visited.member(x)) {
946 continue;
947 } else {
948 visited.push(x);
949 }
950
951 if (x->is_Region()) {
952 for (uint i = 1; i < x->req(); i++) {
953 worklist.push(x->in(i));
954 }
955 } else {
956 worklist.push(x->in(0));
957 // We are looking for the pattern:
958 // /->ThreadLocal
959 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
960 // \->ConI(0)
961 // We want to verify that the If and the LoadB have the same control
962 // See GraphKit::g1_write_barrier_pre()
963 if (x->is_If()) {
964 IfNode *iff = x->as_If();
965 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
966 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
967 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
968 && cmp->in(1)->is_Load()) {
969 LoadNode* load = cmp->in(1)->as_Load();
970 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
971 && load->in(2)->in(3)->is_Con()
972 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
973
974 Node* if_ctrl = iff->in(0);
975 Node* load_ctrl = load->in(0);
976
977 if (if_ctrl != load_ctrl) {
978 // Skip possible CProj->NeverBranch in infinite loops
979 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
980 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
981 if_ctrl = if_ctrl->in(0)->in(0);
982 }
983 }
984 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
985 }
986 }
987 }
988 }
989 }
990 }
991 }
992 #endif
993
994 bool G1BarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
995 if (opcode == Op_StoreP) {
996 Node* adr = n->in(MemNode::Address);
997 const Type* adr_type = gvn->type(adr);
998 // Pointer stores in G1 barriers looks like unsafe access.
999 // Ignore such stores to be able scalar replace non-escaping
1000 // allocations.
1001 if (adr_type->isa_rawptr() && adr->is_AddP()) {
1002 Node* base = conn_graph->get_addp_base(adr);
1003 if (base->Opcode() == Op_LoadP &&
1004 base->in(MemNode::Address)->is_AddP()) {
1005 adr = base->in(MemNode::Address);
1006 Node* tls = conn_graph->get_addp_base(adr);
1007 if (tls->Opcode() == Op_ThreadLocal) {
1008 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1009 const int buf_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset());
1010 if (offs == buf_offset) {
1011 return true; // G1 pre barrier previous oop value store.
1012 }
1013 if (offs == in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset())) {
1014 return true; // G1 post barrier card address store.
1015 }
1016 }
1017 }
1018 }
1019 }
1020 return false;
1021 }