1 /*
2 * Copyright (c) 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/g1/c2/g1BarrierSetC2.hpp"
27 #include "gc/g1/g1BarrierSet.hpp"
28 #include "gc/g1/g1BarrierSetRuntime.hpp"
29 #include "gc/g1/g1CardTable.hpp"
30 #include "gc/g1/g1ThreadLocalData.hpp"
31 #include "gc/g1/heapRegion.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/compile.hpp"
34 #include "opto/graphKit.hpp"
35 #include "opto/idealKit.hpp"
36 #include "opto/macro.hpp"
37 #include "opto/rootnode.hpp"
38 #include "opto/type.hpp"
39 #include "utilities/macros.hpp"
40
41 const TypeFunc *G1BarrierSetC2::write_ref_field_pre_entry_Type() {
42 const Type **fields = TypeTuple::fields(2);
43 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
44 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
45 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
46
47 // create result type (range)
48 fields = TypeTuple::fields(0);
49 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
50
51 return TypeFunc::make(domain, range);
52 }
53
54 const TypeFunc *G1BarrierSetC2::write_ref_field_post_entry_Type() {
55 const Type **fields = TypeTuple::fields(2);
56 fields[TypeFunc::Parms+0] = TypeRawPtr::NOTNULL; // Card addr
57 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
58 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
59
60 // create result type (range)
61 fields = TypeTuple::fields(0);
62 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms, fields);
63
64 return TypeFunc::make(domain, range);
65 }
66
67 #define __ ideal.
68 /*
69 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
70 * required by SATB to make sure all objects live at the start of the
71 * marking are kept alive, all reference updates need to any previous
72 * reference stored before writing.
73 *
74 * If the previous value is NULL there is no need to save the old value.
75 * References that are NULL are filtered during runtime by the barrier
76 * code to avoid unnecessary queuing.
77 *
78 * However in the case of newly allocated objects it might be possible to
79 * prove that the reference about to be overwritten is NULL during compile
80 * time and avoid adding the barrier code completely.
81 *
82 * The compiler needs to determine that the object in which a field is about
83 * to be written is newly allocated, and that no prior store to the same field
84 * has happened since the allocation.
85 *
86 * Returns true if the pre-barrier can be removed
87 */
88 bool G1BarrierSetC2::g1_can_remove_pre_barrier(GraphKit* kit,
89 PhaseTransform* phase,
90 Node* adr,
91 BasicType bt,
92 uint adr_idx) const {
93 intptr_t offset = 0;
94 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
95 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
96
97 if (offset == Type::OffsetBot) {
98 return false; // cannot unalias unless there are precise offsets
99 }
100
101 if (alloc == NULL) {
102 return false; // No allocation found
103 }
104
105 intptr_t size_in_bytes = type2aelembytes(bt);
106
107 Node* mem = kit->memory(adr_idx); // start searching here...
108
109 for (int cnt = 0; cnt < 50; cnt++) {
110
111 if (mem->is_Store()) {
112
113 Node* st_adr = mem->in(MemNode::Address);
114 intptr_t st_offset = 0;
115 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
116
117 if (st_base == NULL) {
118 break; // inscrutable pointer
119 }
120
121 // Break we have found a store with same base and offset as ours so break
122 if (st_base == base && st_offset == offset) {
123 break;
124 }
125
126 if (st_offset != offset && st_offset != Type::OffsetBot) {
127 const int MAX_STORE = BytesPerLong;
128 if (st_offset >= offset + size_in_bytes ||
129 st_offset <= offset - MAX_STORE ||
130 st_offset <= offset - mem->as_Store()->memory_size()) {
131 // Success: The offsets are provably independent.
132 // (You may ask, why not just test st_offset != offset and be done?
133 // The answer is that stores of different sizes can co-exist
134 // in the same sequence of RawMem effects. We sometimes initialize
135 // a whole 'tile' of array elements with a single jint or jlong.)
136 mem = mem->in(MemNode::Memory);
137 continue; // advance through independent store memory
138 }
139 }
140
141 if (st_base != base
142 && MemNode::detect_ptr_independence(base, alloc, st_base,
143 AllocateNode::Ideal_allocation(st_base, phase),
144 phase)) {
145 // Success: The bases are provably independent.
146 mem = mem->in(MemNode::Memory);
147 continue; // advance through independent store memory
148 }
149 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
150
151 InitializeNode* st_init = mem->in(0)->as_Initialize();
152 AllocateNode* st_alloc = st_init->allocation();
153
154 // Make sure that we are looking at the same allocation site.
155 // The alloc variable is guaranteed to not be null here from earlier check.
156 if (alloc == st_alloc) {
157 // Check that the initialization is storing NULL so that no previous store
158 // has been moved up and directly write a reference
159 Node* captured_store = st_init->find_captured_store(offset,
160 type2aelembytes(T_OBJECT),
161 phase);
162 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
163 return true;
164 }
165 }
166 }
167
168 // Unless there is an explicit 'continue', we must bail out here,
169 // because 'mem' is an inscrutable memory state (e.g., a call).
170 break;
171 }
172
173 return false;
174 }
175
176 // G1 pre/post barriers
177 void G1BarrierSetC2::pre_barrier(GraphKit* kit,
178 bool do_load,
179 Node* ctl,
180 Node* obj,
181 Node* adr,
182 uint alias_idx,
183 Node* val,
184 const TypeOopPtr* val_type,
185 Node* pre_val,
186 BasicType bt) const {
187 // Some sanity checks
188 // Note: val is unused in this routine.
189
190 if (do_load) {
191 // We need to generate the load of the previous value
192 assert(obj != NULL, "must have a base");
193 assert(adr != NULL, "where are loading from?");
194 assert(pre_val == NULL, "loaded already?");
195 assert(val_type != NULL, "need a type");
196
197 if (use_ReduceInitialCardMarks()
198 && g1_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
199 return;
200 }
201
202 } else {
203 // In this case both val_type and alias_idx are unused.
204 assert(pre_val != NULL, "must be loaded already");
205 // Nothing to be done if pre_val is null.
206 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
207 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
208 }
209 assert(bt == T_OBJECT, "or we shouldn't be here");
210
211 IdealKit ideal(kit, true);
212
213 Node* tls = __ thread(); // ThreadLocalStorage
214
215 Node* no_base = __ top();
216 Node* zero = __ ConI(0);
217 Node* zeroX = __ ConX(0);
218
219 float likely = PROB_LIKELY(0.999);
220 float unlikely = PROB_UNLIKELY(0.999);
221
222 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
223 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
224
225 // Offsets into the thread
226 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
227 const int index_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset());
228 const int buffer_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset());
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 = write_ref_field_pre_entry_Type();
270 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), "write_ref_field_pre_entry", 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 G1BarrierSetC2::g1_can_remove_post_barrier(GraphKit* kit,
309 PhaseTransform* phase, Node* store,
310 Node* adr) const {
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 G1BarrierSetC2::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) const {
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, G1BarrierSetRuntime::write_ref_field_post_entry), "write_ref_field_post_entry", card_adr, __ thread());
370 } __ end_if();
371
372 }
373
374 void G1BarrierSetC2::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) const {
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 = write_ref_field_post_entry_Type();
425
426 // Offsets into the thread
427 const int index_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset());
428 const int buffer_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset());
429
430 // Pointers into the thread
431
432 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
433 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
434
435 // Now some values
436 // Use ctrl to avoid hoisting these values past a safepoint, which could
437 // potentially reset these fields in the JavaThread.
438 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
439 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
440
441 // Convert the store obj pointer to an int prior to doing math on it
442 // Must use ctrl to prevent "integerized oop" existing across safepoint
443 Node* cast = __ CastPX(__ ctrl(), adr);
444
445 // Divide pointer by card size
446 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
447
448 // Combine card table base and card offset
449 Node* card_adr = __ AddP(no_base, byte_map_base_node(kit), card_offset );
450
451 // If we know the value being stored does it cross regions?
452
453 if (val != NULL) {
454 // Does the store cause us to cross regions?
455
456 // Should be able to do an unsigned compare of region_size instead of
457 // and extra shift. Do we have an unsigned compare??
458 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
459 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
460
461 // if (xor_res == 0) same region so skip
462 __ if_then(xor_res, BoolTest::ne, zeroX); {
463
464 // No barrier if we are storing a NULL
465 __ if_then(val, BoolTest::ne, kit->null(), unlikely); {
466
467 // Ok must mark the card if not already dirty
468
469 // load the original value of the card
470 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
471
472 __ if_then(card_val, BoolTest::ne, young_card); {
473 kit->sync_kit(ideal);
474 kit->insert_mem_bar(Op_MemBarVolatile, oop_store);
475 __ sync_kit(kit);
476
477 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
478 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
479 g1_mark_card(kit, ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
480 } __ end_if();
481 } __ end_if();
482 } __ end_if();
483 } __ end_if();
484 } else {
485 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
486 // We don't need a barrier here if the destination is a newly allocated object
487 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
488 // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()).
489 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
490 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
491 __ if_then(card_val, BoolTest::ne, young_card); {
492 g1_mark_card(kit, ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
493 } __ end_if();
494 }
495
496 // Final sync IdealKit and GraphKit.
497 kit->final_sync(ideal);
498 }
499
500 // Helper that guards and inserts a pre-barrier.
501 void G1BarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
502 Node* pre_val, bool need_mem_bar) const {
503 // We could be accessing the referent field of a reference object. If so, when G1
504 // is enabled, we need to log the value in the referent field in an SATB buffer.
505 // This routine performs some compile time filters and generates suitable
506 // runtime filters that guard the pre-barrier code.
507 // Also add memory barrier for non volatile load from the referent field
508 // to prevent commoning of loads across safepoint.
509
510 // Some compile time checks.
511
512 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
513 const TypeX* otype = offset->find_intptr_t_type();
514 if (otype != NULL && otype->is_con() &&
515 otype->get_con() != java_lang_ref_Reference::referent_offset) {
516 // Constant offset but not the reference_offset so just return
517 return;
518 }
519
520 // We only need to generate the runtime guards for instances.
521 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
522 if (btype != NULL) {
523 if (btype->isa_aryptr()) {
524 // Array type so nothing to do
525 return;
526 }
527
528 const TypeInstPtr* itype = btype->isa_instptr();
529 if (itype != NULL) {
530 // Can the klass of base_oop be statically determined to be
531 // _not_ a sub-class of Reference and _not_ Object?
532 ciKlass* klass = itype->klass();
533 if ( klass->is_loaded() &&
534 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
535 !kit->env()->Object_klass()->is_subtype_of(klass)) {
536 return;
537 }
538 }
539 }
540
541 // The compile time filters did not reject base_oop/offset so
542 // we need to generate the following runtime filters
543 //
544 // if (offset == java_lang_ref_Reference::_reference_offset) {
545 // if (instance_of(base, java.lang.ref.Reference)) {
546 // pre_barrier(_, pre_val, ...);
547 // }
548 // }
549
550 float likely = PROB_LIKELY( 0.999);
551 float unlikely = PROB_UNLIKELY(0.999);
552
553 IdealKit ideal(kit);
554
555 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
556
557 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
558 // Update graphKit memory and control from IdealKit.
559 kit->sync_kit(ideal);
560
561 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
562 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
563
564 // Update IdealKit memory and control from graphKit.
565 __ sync_kit(kit);
566
567 Node* one = __ ConI(1);
568 // is_instof == 0 if base_oop == NULL
569 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
570
571 // Update graphKit from IdeakKit.
572 kit->sync_kit(ideal);
573
574 // Use the pre-barrier to record the value in the referent field
575 pre_barrier(kit, false /* do_load */,
576 __ ctrl(),
577 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
578 pre_val /* pre_val */,
579 T_OBJECT);
580 if (need_mem_bar) {
581 // Add memory barrier to prevent commoning reads from this field
582 // across safepoint since GC can change its value.
583 kit->insert_mem_bar(Op_MemBarCPUOrder);
584 }
585 // Update IdealKit from graphKit.
586 __ sync_kit(kit);
587
588 } __ end_if(); // _ref_type != ref_none
589 } __ end_if(); // offset == referent_offset
590
591 // Final sync IdealKit and GraphKit.
592 kit->final_sync(ideal);
593 }
594
595 #undef __
596
597 Node* G1BarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
598 DecoratorSet decorators = access.decorators();
599 GraphKit* kit = access.kit();
600
601 Node* adr = access.addr().node();
602 Node* obj = access.base();
603
604 bool mismatched = (decorators & C2_MISMATCHED) != 0;
605 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
606 bool in_heap = (decorators & IN_HEAP) != 0;
607 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
608 bool is_unordered = (decorators & MO_UNORDERED) != 0;
609 bool need_cpu_mem_bar = !is_unordered || mismatched || !in_heap;
610
611 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : kit->top();
612 Node* load = CardTableBarrierSetC2::load_at_resolved(access, val_type);
613
614 // If we are reading the value of the referent field of a Reference
615 // object (either by using Unsafe directly or through reflection)
616 // then, if G1 is enabled, we need to record the referent in an
617 // SATB log buffer using the pre-barrier mechanism.
618 // Also we need to add memory barrier to prevent commoning reads
619 // from this field across safepoint since GC can change its value.
620 bool need_read_barrier = in_heap && (on_weak ||
621 (unknown && offset != kit->top() && obj != kit->top()));
622
623 if (!access.is_oop() || !need_read_barrier) {
624 return load;
625 }
626
627 if (on_weak) {
628 // Use the pre-barrier to record the value in the referent field
629 pre_barrier(kit, false /* do_load */,
630 kit->control(),
631 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
632 load /* pre_val */, T_OBJECT);
633 // Add memory barrier to prevent commoning reads from this field
634 // across safepoint since GC can change its value.
635 kit->insert_mem_bar(Op_MemBarCPUOrder);
636 } else if (unknown) {
637 // We do not require a mem bar inside pre_barrier if need_mem_bar
638 // is set: the barriers would be emitted by us.
639 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
640 }
641
642 return load;
643 }
644
645 bool G1BarrierSetC2::is_gc_barrier_node(Node* node) const {
646 if (CardTableBarrierSetC2::is_gc_barrier_node(node)) {
647 return true;
648 }
649 if (node->Opcode() != Op_CallLeaf) {
650 return false;
651 }
652 CallLeafNode *call = node->as_CallLeaf();
653 if (call->_name == NULL) {
654 return false;
655 }
656
657 return strcmp(call->_name, "write_ref_field_pre_entry") == 0 || strcmp(call->_name, "write_ref_field_post_entry") == 0;
658 }
659
660 void G1BarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const {
661 assert(node->Opcode() == Op_CastP2X, "ConvP2XNode required");
662 assert(node->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
663 // It could be only one user, URShift node, in Object.clone() intrinsic
664 // but the new allocation is passed to arraycopy stub and it could not
665 // be scalar replaced. So we don't check the case.
666
667 // An other case of only one user (Xor) is when the value check for NULL
668 // in G1 post barrier is folded after CCP so the code which used URShift
669 // is removed.
670
671 // Take Region node before eliminating post barrier since it also
672 // eliminates CastP2X node when it has only one user.
673 Node* this_region = node->in(0);
674 assert(this_region != NULL, "");
675
676 // Remove G1 post barrier.
677
678 // Search for CastP2X->Xor->URShift->Cmp path which
679 // checks if the store done to a different from the value's region.
680 // And replace Cmp with #0 (false) to collapse G1 post barrier.
681 Node* xorx = node->find_out_with(Op_XorX);
682 if (xorx != NULL) {
683 Node* shift = xorx->unique_out();
684 Node* cmpx = shift->unique_out();
685 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
686 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
687 "missing region check in G1 post barrier");
688 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
689
690 // Remove G1 pre barrier.
691
692 // Search "if (marking != 0)" check and set it to "false".
693 // There is no G1 pre barrier if previous stored value is NULL
694 // (for example, after initialization).
695 if (this_region->is_Region() && this_region->req() == 3) {
696 int ind = 1;
697 if (!this_region->in(ind)->is_IfFalse()) {
698 ind = 2;
699 }
700 if (this_region->in(ind)->is_IfFalse() &&
701 this_region->in(ind)->in(0)->Opcode() == Op_If) {
702 Node* bol = this_region->in(ind)->in(0)->in(1);
703 assert(bol->is_Bool(), "");
704 cmpx = bol->in(1);
705 if (bol->as_Bool()->_test._test == BoolTest::ne &&
706 cmpx->is_Cmp() && cmpx->in(2) == macro->intcon(0) &&
707 cmpx->in(1)->is_Load()) {
708 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
709 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
710 if (adr->is_AddP() && adr->in(AddPNode::Base) == macro->top() &&
711 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
712 adr->in(AddPNode::Offset) == macro->MakeConX(marking_offset)) {
713 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
714 }
715 }
716 }
717 }
718 } else {
719 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
720 // This is a G1 post barrier emitted by the Object.clone() intrinsic.
721 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card
722 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier.
723 Node* shift = node->find_out_with(Op_URShiftX);
724 assert(shift != NULL, "missing G1 post barrier");
725 Node* addp = shift->unique_out();
726 Node* load = addp->find_out_with(Op_LoadB);
727 assert(load != NULL, "missing G1 post barrier");
728 Node* cmpx = load->unique_out();
729 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
730 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
731 "missing card value check in G1 post barrier");
732 macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
733 // There is no G1 pre barrier in this case
734 }
735 // Now CastP2X can be removed since it is used only on dead path
736 // which currently still alive until igvn optimize it.
737 assert(node->outcnt() == 0 || node->unique_out()->Opcode() == Op_URShiftX, "");
738 macro->replace_node(node, macro->top());
739 }
740
741 Node* G1BarrierSetC2::step_over_gc_barrier(Node* c) const {
742 if (!use_ReduceInitialCardMarks() &&
743 c != NULL && c->is_Region() && c->req() == 3) {
744 for (uint i = 1; i < c->req(); i++) {
745 if (c->in(i) != NULL && c->in(i)->is_Region() &&
746 c->in(i)->req() == 3) {
747 Node* r = c->in(i);
748 for (uint j = 1; j < r->req(); j++) {
749 if (r->in(j) != NULL && r->in(j)->is_Proj() &&
750 r->in(j)->in(0) != NULL &&
751 r->in(j)->in(0)->Opcode() == Op_CallLeaf &&
752 r->in(j)->in(0)->as_Call()->entry_point() == CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_post_entry)) {
753 Node* call = r->in(j)->in(0);
754 c = c->in(i == 1 ? 2 : 1);
755 if (c != NULL) {
756 c = c->in(0);
757 if (c != NULL) {
758 c = c->in(0);
759 assert(call->in(0) == NULL ||
760 call->in(0)->in(0) == NULL ||
761 call->in(0)->in(0)->in(0) == NULL ||
762 call->in(0)->in(0)->in(0)->in(0) == NULL ||
763 call->in(0)->in(0)->in(0)->in(0)->in(0) == NULL ||
764 c == call->in(0)->in(0)->in(0)->in(0)->in(0), "bad barrier shape");
765 return c;
766 }
767 }
768 }
769 }
770 }
771 }
772 }
773 return c;
774 }
775
776 #ifdef ASSERT
777 void G1BarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
778 // Verify G1 pre-barriers
779 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
780
781 ResourceArea *area = Thread::current()->resource_area();
782 Unique_Node_List visited(area);
783 Node_List worklist(area);
784 // We're going to walk control flow backwards starting from the Root
785 worklist.push(compile->root());
786 while (worklist.size() > 0) {
787 Node* x = worklist.pop();
788 if (x == NULL || x == compile->top()) continue;
789 if (visited.member(x)) {
790 continue;
791 } else {
792 visited.push(x);
793 }
794
795 if (x->is_Region()) {
796 for (uint i = 1; i < x->req(); i++) {
797 worklist.push(x->in(i));
798 }
799 } else {
800 worklist.push(x->in(0));
801 // We are looking for the pattern:
802 // /->ThreadLocal
803 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
804 // \->ConI(0)
805 // We want to verify that the If and the LoadB have the same control
806 // See GraphKit::g1_write_barrier_pre()
807 if (x->is_If()) {
808 IfNode *iff = x->as_If();
809 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
810 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
811 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
812 && cmp->in(1)->is_Load()) {
813 LoadNode* load = cmp->in(1)->as_Load();
814 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
815 && load->in(2)->in(3)->is_Con()
816 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
817
818 Node* if_ctrl = iff->in(0);
819 Node* load_ctrl = load->in(0);
820
821 if (if_ctrl != load_ctrl) {
822 // Skip possible CProj->NeverBranch in infinite loops
823 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
824 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
825 if_ctrl = if_ctrl->in(0)->in(0);
826 }
827 }
828 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
829 }
830 }
831 }
832 }
833 }
834 }
835 }
836 #endif