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 "classfile/javaClasses.hpp"
27 #include "gc/g1/c2/g1BarrierSetC2.hpp"
28 #include "gc/g1/g1BarrierSet.hpp"
29 #include "gc/g1/g1BarrierSetRuntime.hpp"
30 #include "gc/g1/g1CardTable.hpp"
31 #include "gc/g1/g1ThreadLocalData.hpp"
32 #include "gc/g1/heapRegion.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/compile.hpp"
35 #include "opto/escape.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/idealKit.hpp"
38 #include "opto/macro.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/type.hpp"
41 #include "utilities/macros.hpp"
42
43 const TypeFunc *G1BarrierSetC2::write_ref_field_pre_entry_Type() {
44 const Type **fields = TypeTuple::fields(2);
45 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
46 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
47 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
48
49 // create result type (range)
50 fields = TypeTuple::fields(0);
51 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
52
53 return TypeFunc::make(domain, range);
54 }
55
56 const TypeFunc *G1BarrierSetC2::write_ref_field_post_entry_Type() {
57 const Type **fields = TypeTuple::fields(2);
58 fields[TypeFunc::Parms+0] = TypeRawPtr::NOTNULL; // Card addr
59 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
60 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
61
62 // create result type (range)
63 fields = TypeTuple::fields(0);
64 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms, fields);
65
66 return TypeFunc::make(domain, range);
67 }
68
69 #define __ ideal.
70 /*
71 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
72 * required by SATB to make sure all objects live at the start of the
73 * marking are kept alive, all reference updates need to any previous
74 * reference stored before writing.
75 *
76 * If the previous value is NULL there is no need to save the old value.
77 * References that are NULL are filtered during runtime by the barrier
78 * code to avoid unnecessary queuing.
79 *
80 * However in the case of newly allocated objects it might be possible to
81 * prove that the reference about to be overwritten is NULL during compile
82 * time and avoid adding the barrier code completely.
83 *
84 * The compiler needs to determine that the object in which a field is about
85 * to be written is newly allocated, and that no prior store to the same field
86 * has happened since the allocation.
87 *
88 * Returns true if the pre-barrier can be removed
89 */
90 bool G1BarrierSetC2::g1_can_remove_pre_barrier(GraphKit* kit,
91 PhaseTransform* phase,
92 Node* adr,
93 BasicType bt,
94 uint adr_idx) const {
95 intptr_t offset = 0;
96 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
97 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
98
99 if (offset == Type::OffsetBot) {
100 return false; // cannot unalias unless there are precise offsets
101 }
102
103 if (alloc == NULL) {
104 return false; // No allocation found
105 }
106
107 intptr_t size_in_bytes = type2aelembytes(bt);
108
109 Node* mem = kit->memory(adr_idx); // start searching here...
110
111 for (int cnt = 0; cnt < 50; cnt++) {
112
113 if (mem->is_Store()) {
114
115 Node* st_adr = mem->in(MemNode::Address);
116 intptr_t st_offset = 0;
117 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
118
119 if (st_base == NULL) {
120 break; // inscrutable pointer
121 }
122
123 // Break we have found a store with same base and offset as ours so break
124 if (st_base == base && st_offset == offset) {
125 break;
126 }
127
128 if (st_offset != offset && st_offset != Type::OffsetBot) {
129 const int MAX_STORE = BytesPerLong;
130 if (st_offset >= offset + size_in_bytes ||
131 st_offset <= offset - MAX_STORE ||
132 st_offset <= offset - mem->as_Store()->memory_size()) {
133 // Success: The offsets are provably independent.
134 // (You may ask, why not just test st_offset != offset and be done?
135 // The answer is that stores of different sizes can co-exist
136 // in the same sequence of RawMem effects. We sometimes initialize
137 // a whole 'tile' of array elements with a single jint or jlong.)
138 mem = mem->in(MemNode::Memory);
139 continue; // advance through independent store memory
140 }
141 }
142
143 if (st_base != base
144 && MemNode::detect_ptr_independence(base, alloc, st_base,
145 AllocateNode::Ideal_allocation(st_base, phase),
146 phase)) {
147 // Success: The bases are provably independent.
148 mem = mem->in(MemNode::Memory);
149 continue; // advance through independent store memory
150 }
151 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
152
153 InitializeNode* st_init = mem->in(0)->as_Initialize();
154 AllocateNode* st_alloc = st_init->allocation();
155
156 // Make sure that we are looking at the same allocation site.
157 // The alloc variable is guaranteed to not be null here from earlier check.
158 if (alloc == st_alloc) {
159 // Check that the initialization is storing NULL so that no previous store
160 // has been moved up and directly write a reference
161 Node* captured_store = st_init->find_captured_store(offset,
162 type2aelembytes(T_OBJECT),
163 phase);
164 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
165 return true;
166 }
167 }
168 }
169
170 // Unless there is an explicit 'continue', we must bail out here,
171 // because 'mem' is an inscrutable memory state (e.g., a call).
172 break;
173 }
174
175 return false;
176 }
177
178 // G1 pre/post barriers
179 void G1BarrierSetC2::pre_barrier(GraphKit* kit,
180 bool do_load,
181 Node* ctl,
182 Node* obj,
183 Node* adr,
184 uint alias_idx,
185 Node* val,
186 const TypeOopPtr* val_type,
187 Node* pre_val,
188 BasicType bt) const {
189 // Some sanity checks
190 // Note: val is unused in this routine.
191
192 if (do_load) {
193 // We need to generate the load of the previous value
194 assert(obj != NULL, "must have a base");
195 assert(adr != NULL, "where are loading from?");
196 assert(pre_val == NULL, "loaded already?");
197 assert(val_type != NULL, "need a type");
198
199 if (use_ReduceInitialCardMarks()
200 && g1_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
201 return;
202 }
203
204 } else {
205 // In this case both val_type and alias_idx are unused.
206 assert(pre_val != NULL, "must be loaded already");
207 // Nothing to be done if pre_val is null.
208 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
209 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
210 }
211 assert(bt == T_OBJECT, "or we shouldn't be here");
212
213 IdealKit ideal(kit, true);
214
215 Node* tls = __ thread(); // ThreadLocalStorage
216
217 Node* no_base = __ top();
218 Node* zero = __ ConI(0);
219 Node* zeroX = __ ConX(0);
220
221 float likely = PROB_LIKELY(0.999);
222 float unlikely = PROB_UNLIKELY(0.999);
223
224 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
225 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
226
227 // Offsets into the thread
228 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
229 const int index_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset());
230 const int buffer_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset());
231
232 // Now the actual pointers into the thread
233 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
234 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
235 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
236
237 // Now some of the values
238 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
239
240 // if (!marking)
241 __ if_then(marking, BoolTest::ne, zero, unlikely); {
242 BasicType index_bt = TypeX_X->basic_type();
243 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
244 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
245
246 if (do_load) {
247 // load original value
248 // alias_idx correct??
249 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
250 }
251
252 // if (pre_val != NULL)
253 __ if_then(pre_val, BoolTest::ne, kit->null()); {
254 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
255
256 // is the queue for this thread full?
257 __ if_then(index, BoolTest::ne, zeroX, likely); {
258
259 // decrement the index
260 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
261
262 // Now get the buffer location we will log the previous value into and store it
263 Node *log_addr = __ AddP(no_base, buffer, next_index);
264 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
265 // update the index
266 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
267
268 } __ else_(); {
269
270 // logging buffer is full, call the runtime
271 const TypeFunc *tf = write_ref_field_pre_entry_Type();
272 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), "write_ref_field_pre_entry", pre_val, tls);
273 } __ end_if(); // (!index)
274 } __ end_if(); // (pre_val != NULL)
275 } __ end_if(); // (!marking)
276
277 // Final sync IdealKit and GraphKit.
278 kit->final_sync(ideal);
279 }
280
281 /*
282 * G1 similar to any GC with a Young Generation requires a way to keep track of
283 * references from Old Generation to Young Generation to make sure all live
284 * objects are found. G1 also requires to keep track of object references
285 * between different regions to enable evacuation of old regions, which is done
286 * as part of mixed collections. References are tracked in remembered sets and
287 * is continuously updated as reference are written to with the help of the
288 * post-barrier.
289 *
290 * To reduce the number of updates to the remembered set the post-barrier
291 * filters updates to fields in objects located in the Young Generation,
292 * the same region as the reference, when the NULL is being written or
293 * if the card is already marked as dirty by an earlier write.
294 *
295 * Under certain circumstances it is possible to avoid generating the
296 * post-barrier completely if it is possible during compile time to prove
297 * the object is newly allocated and that no safepoint exists between the
298 * allocation and the store.
299 *
300 * In the case of slow allocation the allocation code must handle the barrier
301 * as part of the allocation in the case the allocated object is not located
302 * in the nursery; this would happen for humongous objects.
303 *
304 * Returns true if the post barrier can be removed
305 */
306 bool G1BarrierSetC2::g1_can_remove_post_barrier(GraphKit* kit,
307 PhaseTransform* phase, Node* store,
308 Node* adr) const {
309 intptr_t offset = 0;
310 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
311 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
312
313 if (offset == Type::OffsetBot) {
314 return false; // cannot unalias unless there are precise offsets
315 }
316
317 if (alloc == NULL) {
318 return false; // No allocation found
319 }
320
321 // Start search from Store node
322 Node* mem = store->in(MemNode::Control);
323 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
324
325 InitializeNode* st_init = mem->in(0)->as_Initialize();
326 AllocateNode* st_alloc = st_init->allocation();
327
328 // Make sure we are looking at the same allocation
329 if (alloc == st_alloc) {
330 return true;
331 }
332 }
333
334 return false;
335 }
336
337 //
338 // Update the card table and add card address to the queue
339 //
340 void G1BarrierSetC2::g1_mark_card(GraphKit* kit,
341 IdealKit& ideal,
342 Node* card_adr,
343 Node* oop_store,
344 uint oop_alias_idx,
345 Node* index,
346 Node* index_adr,
347 Node* buffer,
348 const TypeFunc* tf) const {
349 Node* zero = __ ConI(0);
350 Node* zeroX = __ ConX(0);
351 Node* no_base = __ top();
352 BasicType card_bt = T_BYTE;
353 // Smash zero into card. MUST BE ORDERED WRT TO STORE
354 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
355
356 // Now do the queue work
357 __ if_then(index, BoolTest::ne, zeroX); {
358
359 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
360 Node* log_addr = __ AddP(no_base, buffer, next_index);
361
362 // Order, see storeCM.
363 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
364 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
365
366 } __ else_(); {
367 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_post_entry), "write_ref_field_post_entry", card_adr, __ thread());
368 } __ end_if();
369
370 }
371
372 void G1BarrierSetC2::post_barrier(GraphKit* kit,
373 Node* ctl,
374 Node* oop_store,
375 Node* obj,
376 Node* adr,
377 uint alias_idx,
378 Node* val,
379 BasicType bt,
380 bool use_precise) const {
381 // If we are writing a NULL then we need no post barrier
382
383 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
384 // Must be NULL
385 const Type* t = val->bottom_type();
386 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
387 // No post barrier if writing NULLx
388 return;
389 }
390
391 if (use_ReduceInitialCardMarks() && obj == kit->just_allocated_object(kit->control())) {
392 // We can skip marks on a freshly-allocated object in Eden.
393 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
394 // That routine informs GC to take appropriate compensating steps,
395 // upon a slow-path allocation, so as to make this card-mark
396 // elision safe.
397 return;
398 }
399
400 if (use_ReduceInitialCardMarks()
401 && g1_can_remove_post_barrier(kit, &kit->gvn(), oop_store, adr)) {
402 return;
403 }
404
405 if (!use_precise) {
406 // All card marks for a (non-array) instance are in one place:
407 adr = obj;
408 }
409 // (Else it's an array (or unknown), and we want more precise card marks.)
410 assert(adr != NULL, "");
411
412 IdealKit ideal(kit, true);
413
414 Node* tls = __ thread(); // ThreadLocalStorage
415
416 Node* no_base = __ top();
417 float likely = PROB_LIKELY_MAG(3);
418 float unlikely = PROB_UNLIKELY_MAG(3);
419 Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val());
420 Node* dirty_card = __ ConI((jint)G1CardTable::dirty_card_val());
421 Node* zeroX = __ ConX(0);
422
423 const TypeFunc *tf = write_ref_field_post_entry_Type();
424
425 // Offsets into the thread
426 const int index_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset());
427 const int buffer_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset());
428
429 // Pointers into the thread
430
431 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
432 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
433
434 // Now some values
435 // Use ctrl to avoid hoisting these values past a safepoint, which could
436 // potentially reset these fields in the JavaThread.
437 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
438 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
439
440 // Convert the store obj pointer to an int prior to doing math on it
441 // Must use ctrl to prevent "integerized oop" existing across safepoint
442 Node* cast = __ CastPX(__ ctrl(), adr);
443
444 // Divide pointer by card size
445 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift()) );
446
447 // Combine card table base and card offset
448 Node* card_adr = __ AddP(no_base, byte_map_base_node(kit), card_offset );
449
450 // If we know the value being stored does it cross regions?
451
452 if (val != NULL) {
453 // Does the store cause us to cross regions?
454
455 // Should be able to do an unsigned compare of region_size instead of
456 // and extra shift. Do we have an unsigned compare??
457 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
458 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
459
460 // if (xor_res == 0) same region so skip
461 __ if_then(xor_res, BoolTest::ne, zeroX, likely); {
462
463 // No barrier if we are storing a NULL
464 __ if_then(val, BoolTest::ne, kit->null(), likely); {
465
466 // Ok must mark the card if not already dirty
467
468 // load the original value of the card
469 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
470
471 __ if_then(card_val, BoolTest::ne, young_card, unlikely); {
472 kit->sync_kit(ideal);
473 kit->insert_mem_bar(Op_MemBarVolatile, oop_store);
474 __ sync_kit(kit);
475
476 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
477 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
478 g1_mark_card(kit, ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
479 } __ end_if();
480 } __ end_if();
481 } __ end_if();
482 } __ end_if();
483 } else {
484 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
485 // We don't need a barrier here if the destination is a newly allocated object
486 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
487 // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()).
488 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
489 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
490 __ if_then(card_val, BoolTest::ne, young_card); {
491 g1_mark_card(kit, ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
492 } __ end_if();
493 }
494
495 // Final sync IdealKit and GraphKit.
496 kit->final_sync(ideal);
497 }
498
499 // Helper that guards and inserts a pre-barrier.
500 void G1BarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
501 Node* pre_val, bool need_mem_bar) const {
502 // We could be accessing the referent field of a reference object. If so, when G1
503 // is enabled, we need to log the value in the referent field in an SATB buffer.
504 // This routine performs some compile time filters and generates suitable
505 // runtime filters that guard the pre-barrier code.
506 // Also add memory barrier for non volatile load from the referent field
507 // to prevent commoning of loads across safepoint.
508
509 // Some compile time checks.
510
511 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
512 const TypeX* otype = offset->find_intptr_t_type();
513 if (otype != NULL && otype->is_con() &&
514 otype->get_con() != java_lang_ref_Reference::referent_offset()) {
515 // Constant offset but not the reference_offset so just return
516 return;
517 }
518
519 // We only need to generate the runtime guards for instances.
520 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
521 if (btype != NULL) {
522 if (btype->isa_aryptr()) {
523 // Array type so nothing to do
524 return;
525 }
526
527 const TypeInstPtr* itype = btype->isa_instptr();
528 if (itype != NULL) {
529 // Can the klass of base_oop be statically determined to be
530 // _not_ a sub-class of Reference and _not_ Object?
531 ciKlass* klass = itype->instance_klass();
532 if (klass->is_loaded() &&
533 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
534 !kit->env()->Object_klass()->is_subtype_of(klass)) {
535 return;
536 }
537 }
538 }
539
540 // The compile time filters did not reject base_oop/offset so
541 // we need to generate the following runtime filters
542 //
543 // if (offset == java_lang_ref_Reference::_reference_offset) {
544 // if (instance_of(base, java.lang.ref.Reference)) {
545 // pre_barrier(_, pre_val, ...);
546 // }
547 // }
548
549 float likely = PROB_LIKELY( 0.999);
550 float unlikely = PROB_UNLIKELY(0.999);
551
552 IdealKit ideal(kit);
553
554 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset());
555
556 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
557 // Update graphKit memory and control from IdealKit.
558 kit->sync_kit(ideal);
559
560 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
561 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
562
563 // Update IdealKit memory and control from graphKit.
564 __ sync_kit(kit);
565
566 Node* one = __ ConI(1);
567 // is_instof == 0 if base_oop == NULL
568 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
569
570 // Update graphKit from IdeakKit.
571 kit->sync_kit(ideal);
572
573 // Use the pre-barrier to record the value in the referent field
574 pre_barrier(kit, false /* do_load */,
575 __ ctrl(),
576 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
577 pre_val /* pre_val */,
578 T_OBJECT);
579 if (need_mem_bar) {
580 // Add memory barrier to prevent commoning reads from this field
581 // across safepoint since GC can change its value.
582 kit->insert_mem_bar(Op_MemBarCPUOrder);
583 }
584 // Update IdealKit from graphKit.
585 __ sync_kit(kit);
586
587 } __ end_if(); // _ref_type != ref_none
588 } __ end_if(); // offset == referent_offset
589
590 // Final sync IdealKit and GraphKit.
591 kit->final_sync(ideal);
592 }
593
594 #undef __
595
596 Node* G1BarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
597 DecoratorSet decorators = access.decorators();
598 Node* adr = access.addr().node();
599 Node* obj = access.base();
600
601 bool anonymous = (decorators & C2_UNSAFE_ACCESS) != 0;
602 bool mismatched = (decorators & C2_MISMATCHED) != 0;
603 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
604 bool in_heap = (decorators & IN_HEAP) != 0;
605 bool in_native = (decorators & IN_NATIVE) != 0;
606 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
607 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
608 bool is_unordered = (decorators & MO_UNORDERED) != 0;
609 bool no_keepalive = (decorators & AS_NO_KEEPALIVE) != 0;
610 bool is_mixed = !in_heap && !in_native;
611 bool need_cpu_mem_bar = !is_unordered || mismatched || is_mixed;
612
613 Node* top = Compile::current()->top();
614 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
615 Node* load = CardTableBarrierSetC2::load_at_resolved(access, val_type);
616
617 // If we are reading the value of the referent field of a Reference
618 // object (either by using Unsafe directly or through reflection)
619 // then, if G1 is enabled, we need to record the referent in an
620 // SATB log buffer using the pre-barrier mechanism.
621 // Also we need to add memory barrier to prevent commoning reads
622 // from this field across safepoint since GC can change its value.
623 bool need_read_barrier = (((on_weak || on_phantom) && !no_keepalive) ||
624 (in_heap && unknown && offset != top && obj != top));
625
626 if (!access.is_oop() || !need_read_barrier) {
627 return load;
628 }
629
630 assert(access.is_parse_access(), "entry not supported at optimization time");
631 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
632 GraphKit* kit = parse_access.kit();
633
634 if (on_weak || on_phantom) {
635 // Use the pre-barrier to record the value in the referent field
636 pre_barrier(kit, false /* do_load */,
637 kit->control(),
638 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
639 load /* pre_val */, T_OBJECT);
640 // Add memory barrier to prevent commoning reads from this field
641 // across safepoint since GC can change its value.
642 kit->insert_mem_bar(Op_MemBarCPUOrder);
643 } else if (unknown) {
644 // We do not require a mem bar inside pre_barrier if need_mem_bar
645 // is set: the barriers would be emitted by us.
646 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
647 }
648
649 return load;
650 }
651
652 bool G1BarrierSetC2::is_gc_barrier_node(Node* node) const {
653 if (CardTableBarrierSetC2::is_gc_barrier_node(node)) {
654 return true;
655 }
656 if (node->Opcode() != Op_CallLeaf) {
657 return false;
658 }
659 CallLeafNode *call = node->as_CallLeaf();
660 if (call->_name == NULL) {
661 return false;
662 }
663
664 return strcmp(call->_name, "write_ref_field_pre_entry") == 0 || strcmp(call->_name, "write_ref_field_post_entry") == 0;
665 }
666
667 void G1BarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const {
668 assert(node->Opcode() == Op_CastP2X, "ConvP2XNode required");
669 assert(node->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
670 // It could be only one user, URShift node, in Object.clone() intrinsic
671 // but the new allocation is passed to arraycopy stub and it could not
672 // be scalar replaced. So we don't check the case.
673
674 // An other case of only one user (Xor) is when the value check for NULL
675 // in G1 post barrier is folded after CCP so the code which used URShift
676 // is removed.
677
678 // Take Region node before eliminating post barrier since it also
679 // eliminates CastP2X node when it has only one user.
680 Node* this_region = node->in(0);
681 assert(this_region != NULL, "");
682
683 // Remove G1 post barrier.
684
685 // Search for CastP2X->Xor->URShift->Cmp path which
686 // checks if the store done to a different from the value's region.
687 // And replace Cmp with #0 (false) to collapse G1 post barrier.
688 Node* xorx = node->find_out_with(Op_XorX);
689 if (xorx != NULL) {
690 Node* shift = xorx->unique_out();
691 Node* cmpx = shift->unique_out();
692 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
693 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
694 "missing region check in G1 post barrier");
695 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
696
697 // Remove G1 pre barrier.
698
699 // Search "if (marking != 0)" check and set it to "false".
700 // There is no G1 pre barrier if previous stored value is NULL
701 // (for example, after initialization).
702 if (this_region->is_Region() && this_region->req() == 3) {
703 int ind = 1;
704 if (!this_region->in(ind)->is_IfFalse()) {
705 ind = 2;
706 }
707 if (this_region->in(ind)->is_IfFalse() &&
708 this_region->in(ind)->in(0)->Opcode() == Op_If) {
709 Node* bol = this_region->in(ind)->in(0)->in(1);
710 assert(bol->is_Bool(), "");
711 cmpx = bol->in(1);
712 if (bol->as_Bool()->_test._test == BoolTest::ne &&
713 cmpx->is_Cmp() && cmpx->in(2) == macro->intcon(0) &&
714 cmpx->in(1)->is_Load()) {
715 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
716 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
717 if (adr->is_AddP() && adr->in(AddPNode::Base) == macro->top() &&
718 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
719 adr->in(AddPNode::Offset) == macro->MakeConX(marking_offset)) {
720 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
721 }
722 }
723 }
724 }
725 } else {
726 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
727 // This is a G1 post barrier emitted by the Object.clone() intrinsic.
728 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card
729 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier.
730 Node* shift = node->find_out_with(Op_URShiftX);
731 assert(shift != NULL, "missing G1 post barrier");
732 Node* addp = shift->unique_out();
733 Node* load = addp->find_out_with(Op_LoadB);
734 assert(load != NULL, "missing G1 post barrier");
735 Node* cmpx = load->unique_out();
736 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
737 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
738 "missing card value check in G1 post barrier");
739 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
740 // There is no G1 pre barrier in this case
741 }
742 // Now CastP2X can be removed since it is used only on dead path
743 // which currently still alive until igvn optimize it.
744 assert(node->outcnt() == 0 || node->unique_out()->Opcode() == Op_URShiftX, "");
745 macro->replace_node(node, macro->top());
746 }
747
748 Node* G1BarrierSetC2::step_over_gc_barrier(Node* c) const {
749 if (!use_ReduceInitialCardMarks() &&
750 c != NULL && c->is_Region() && c->req() == 3) {
751 for (uint i = 1; i < c->req(); i++) {
752 if (c->in(i) != NULL && c->in(i)->is_Region() &&
753 c->in(i)->req() == 3) {
754 Node* r = c->in(i);
755 for (uint j = 1; j < r->req(); j++) {
756 if (r->in(j) != NULL && r->in(j)->is_Proj() &&
757 r->in(j)->in(0) != NULL &&
758 r->in(j)->in(0)->Opcode() == Op_CallLeaf &&
759 r->in(j)->in(0)->as_Call()->entry_point() == CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_post_entry)) {
760 Node* call = r->in(j)->in(0);
761 c = c->in(i == 1 ? 2 : 1);
762 if (c != NULL && c->Opcode() != Op_Parm) {
763 c = c->in(0);
764 if (c != NULL) {
765 c = c->in(0);
766 assert(call->in(0) == NULL ||
767 call->in(0)->in(0) == NULL ||
768 call->in(0)->in(0)->in(0) == NULL ||
769 call->in(0)->in(0)->in(0)->in(0) == NULL ||
770 call->in(0)->in(0)->in(0)->in(0)->in(0) == NULL ||
771 c == call->in(0)->in(0)->in(0)->in(0)->in(0), "bad barrier shape");
772 return c;
773 }
774 }
775 }
776 }
777 }
778 }
779 }
780 return c;
781 }
782
783 #ifdef ASSERT
784 void G1BarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
785 if (phase != BarrierSetC2::BeforeCodeGen) {
786 return;
787 }
788 // Verify G1 pre-barriers
789 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
790
791 Unique_Node_List visited;
792 Node_List worklist;
793 // We're going to walk control flow backwards starting from the Root
794 worklist.push(compile->root());
795 while (worklist.size() > 0) {
796 Node* x = worklist.pop();
797 if (x == NULL || x == compile->top()) continue;
798 if (visited.member(x)) {
799 continue;
800 } else {
801 visited.push(x);
802 }
803
804 if (x->is_Region()) {
805 for (uint i = 1; i < x->req(); i++) {
806 worklist.push(x->in(i));
807 }
808 } else {
809 worklist.push(x->in(0));
810 // We are looking for the pattern:
811 // /->ThreadLocal
812 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
813 // \->ConI(0)
814 // We want to verify that the If and the LoadB have the same control
815 // See GraphKit::g1_write_barrier_pre()
816 if (x->is_If()) {
817 IfNode *iff = x->as_If();
818 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
819 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
820 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
821 && cmp->in(1)->is_Load()) {
822 LoadNode* load = cmp->in(1)->as_Load();
823 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
824 && load->in(2)->in(3)->is_Con()
825 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
826
827 Node* if_ctrl = iff->in(0);
828 Node* load_ctrl = load->in(0);
829
830 if (if_ctrl != load_ctrl) {
831 // Skip possible CProj->NeverBranch in infinite loops
832 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
833 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
834 if_ctrl = if_ctrl->in(0)->in(0);
835 }
836 }
837 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
838 }
839 }
840 }
841 }
842 }
843 }
844 }
845 #endif
846
847 bool G1BarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
848 if (opcode == Op_StoreP) {
849 Node* adr = n->in(MemNode::Address);
850 const Type* adr_type = gvn->type(adr);
851 // Pointer stores in G1 barriers looks like unsafe access.
852 // Ignore such stores to be able scalar replace non-escaping
853 // allocations.
854 if (adr_type->isa_rawptr() && adr->is_AddP()) {
855 Node* base = conn_graph->get_addp_base(adr);
856 if (base->Opcode() == Op_LoadP &&
857 base->in(MemNode::Address)->is_AddP()) {
858 adr = base->in(MemNode::Address);
859 Node* tls = conn_graph->get_addp_base(adr);
860 if (tls->Opcode() == Op_ThreadLocal) {
861 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
862 const int buf_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset());
863 if (offs == buf_offset) {
864 return true; // G1 pre barrier previous oop value store.
865 }
866 if (offs == in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset())) {
867 return true; // G1 post barrier card address store.
868 }
869 }
870 }
871 }
872 }
873 return false;
874 }