1 /*
2 * Copyright (c) 2018, 2023, Red Hat, Inc. All rights reserved.
3 * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27 #include "classfile/javaClasses.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shenandoah/shenandoahBarrierSet.hpp"
30 #include "gc/shenandoah/shenandoahCardTable.hpp"
31 #include "gc/shenandoah/shenandoahForwarding.hpp"
32 #include "gc/shenandoah/shenandoahHeap.hpp"
33 #include "gc/shenandoah/shenandoahRuntime.hpp"
34 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
35 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
36 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
37 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
38 #include "opto/arraycopynode.hpp"
39 #include "opto/escape.hpp"
40 #include "opto/graphKit.hpp"
41 #include "opto/idealKit.hpp"
42 #include "opto/macro.hpp"
43 #include "opto/movenode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/rootnode.hpp"
46 #include "opto/runtime.hpp"
47
48 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() {
49 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2());
50 }
51
52 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena)
53 : _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8, 0, nullptr)) {
54 }
55
56 int ShenandoahBarrierSetC2State::load_reference_barriers_count() const {
57 return _load_reference_barriers->length();
58 }
59
60 ShenandoahLoadReferenceBarrierNode* ShenandoahBarrierSetC2State::load_reference_barrier(int idx) const {
61 return _load_reference_barriers->at(idx);
62 }
63
64 void ShenandoahBarrierSetC2State::add_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
65 assert(!_load_reference_barriers->contains(n), "duplicate entry in barrier list");
66 _load_reference_barriers->append(n);
67 }
68
69 void ShenandoahBarrierSetC2State::remove_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
70 if (_load_reference_barriers->contains(n)) {
71 _load_reference_barriers->remove(n);
72 }
73 }
74
75 #define __ kit->
76
77 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseValues* phase, Node* adr,
78 BasicType bt, uint adr_idx) const {
79 intptr_t offset = 0;
80 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
81 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
82
83 if (offset == Type::OffsetBot) {
84 return false; // cannot unalias unless there are precise offsets
85 }
86
87 if (alloc == nullptr) {
88 return false; // No allocation found
89 }
90
91 intptr_t size_in_bytes = type2aelembytes(bt);
92
93 Node* mem = __ memory(adr_idx); // start searching here...
94
95 for (int cnt = 0; cnt < 50; cnt++) {
96
97 if (mem->is_Store()) {
98
99 Node* st_adr = mem->in(MemNode::Address);
100 intptr_t st_offset = 0;
101 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
102
103 if (st_base == nullptr) {
104 break; // inscrutable pointer
105 }
106
107 // Break we have found a store with same base and offset as ours so break
108 if (st_base == base && st_offset == offset) {
109 break;
110 }
111
112 if (st_offset != offset && st_offset != Type::OffsetBot) {
113 const int MAX_STORE = BytesPerLong;
114 if (st_offset >= offset + size_in_bytes ||
115 st_offset <= offset - MAX_STORE ||
116 st_offset <= offset - mem->as_Store()->memory_size()) {
117 // Success: The offsets are provably independent.
118 // (You may ask, why not just test st_offset != offset and be done?
119 // The answer is that stores of different sizes can co-exist
120 // in the same sequence of RawMem effects. We sometimes initialize
121 // a whole 'tile' of array elements with a single jint or jlong.)
122 mem = mem->in(MemNode::Memory);
123 continue; // advance through independent store memory
124 }
125 }
126
127 if (st_base != base
128 && MemNode::detect_ptr_independence(base, alloc, st_base,
129 AllocateNode::Ideal_allocation(st_base),
130 phase)) {
131 // Success: The bases are provably independent.
132 mem = mem->in(MemNode::Memory);
133 continue; // advance through independent store memory
134 }
135 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
136
137 InitializeNode* st_init = mem->in(0)->as_Initialize();
138 AllocateNode* st_alloc = st_init->allocation();
139
140 // Make sure that we are looking at the same allocation site.
141 // The alloc variable is guaranteed to not be null here from earlier check.
142 if (alloc == st_alloc) {
143 // Check that the initialization is storing null so that no previous store
144 // has been moved up and directly write a reference
145 Node* captured_store = st_init->find_captured_store(offset,
146 type2aelembytes(T_OBJECT),
147 phase);
148 if (captured_store == nullptr || captured_store == st_init->zero_memory()) {
149 return true;
150 }
151 }
152 }
153
154 // Unless there is an explicit 'continue', we must bail out here,
155 // because 'mem' is an inscrutable memory state (e.g., a call).
156 break;
157 }
158
159 return false;
160 }
161
162 #undef __
163 #define __ ideal.
164
165 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit,
166 bool do_load,
167 Node* obj,
168 Node* adr,
169 uint alias_idx,
170 Node* val,
171 const TypeOopPtr* val_type,
172 Node* pre_val,
173 BasicType bt) const {
174 // Some sanity checks
175 // Note: val is unused in this routine.
176
177 if (do_load) {
178 // We need to generate the load of the previous value
179 assert(adr != nullptr, "where are loading from?");
180 assert(pre_val == nullptr, "loaded already?");
181 assert(val_type != nullptr, "need a type");
182
183 if (ReduceInitialCardMarks
184 && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
185 return;
186 }
187
188 } else {
189 // In this case both val_type and alias_idx are unused.
190 assert(pre_val != nullptr, "must be loaded already");
191 // Nothing to be done if pre_val is null.
192 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
193 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
194 }
195 assert(bt == T_OBJECT, "or we shouldn't be here");
196
197 IdealKit ideal(kit, true);
198
199 Node* tls = __ thread(); // ThreadLocalStorage
200
201 Node* no_base = __ top();
202 Node* zero = __ ConI(0);
203 Node* zeroX = __ ConX(0);
204
205 float likely = PROB_LIKELY(0.999);
206 float unlikely = PROB_UNLIKELY(0.999);
207
208 // Offsets into the thread
209 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
210 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
211
212 // Now the actual pointers into the thread
213 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
214 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
215
216 // Now some of the values
217 Node* marking;
218 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())));
219 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
220 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
221 assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape");
222
223 // if (!marking)
224 __ if_then(marking, BoolTest::ne, zero, unlikely); {
225 BasicType index_bt = TypeX_X->basic_type();
226 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading Shenandoah SATBMarkQueue::_index with wrong size.");
227 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
228
229 if (do_load) {
230 // load original value
231 // alias_idx correct??
232 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
233 }
234
235 // if (pre_val != nullptr)
236 __ if_then(pre_val, BoolTest::ne, kit->null()); {
237 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
238
239 // is the queue for this thread full?
240 __ if_then(index, BoolTest::ne, zeroX, likely); {
241
242 // decrement the index
243 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
244
245 // Now get the buffer location we will log the previous value into and store it
246 Node *log_addr = __ AddP(no_base, buffer, next_index);
247 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
248 // update the index
249 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
250
251 } __ else_(); {
252
253 // logging buffer is full, call the runtime
254 const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type();
255 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls);
256 } __ end_if(); // (!index)
257 } __ end_if(); // (pre_val != nullptr)
258 } __ end_if(); // (!marking)
259
260 // Final sync IdealKit and GraphKit.
261 kit->final_sync(ideal);
262
263 if (ShenandoahSATBBarrier && adr != nullptr) {
264 Node* c = kit->control();
265 Node* call = c->in(1)->in(1)->in(1)->in(0);
266 assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected");
267 call->add_req(adr);
268 }
269 }
270
271 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) {
272 return call->is_CallLeaf() &&
273 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry);
274 }
275
276 bool ShenandoahBarrierSetC2::is_shenandoah_lrb_call(Node* call) {
277 if (!call->is_CallLeaf()) {
278 return false;
279 }
280
281 address entry_point = call->as_CallLeaf()->entry_point();
282 return (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong)) ||
283 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow)) ||
284 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak)) ||
285 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow)) ||
286 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom));
287 }
288
289 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseValues* phase, Node* n) {
290 if (n->Opcode() != Op_If) {
291 return false;
292 }
293
294 Node* bol = n->in(1);
295 assert(bol->is_Bool(), "");
296 Node* cmpx = bol->in(1);
297 if (bol->as_Bool()->_test._test == BoolTest::ne &&
298 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
299 is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
300 cmpx->in(1)->in(2)->is_Con() &&
301 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
302 return true;
303 }
304
305 return false;
306 }
307
308 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
309 if (!n->is_Load()) return false;
310 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
311 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
312 && n->in(2)->in(3)->is_Con()
313 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
314 }
315
316 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
317 bool do_load,
318 Node* obj,
319 Node* adr,
320 uint alias_idx,
321 Node* val,
322 const TypeOopPtr* val_type,
323 Node* pre_val,
324 BasicType bt) const {
325 if (ShenandoahSATBBarrier) {
326 IdealKit ideal(kit);
327 kit->sync_kit(ideal);
328
329 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
330
331 ideal.sync_kit(kit);
332 kit->final_sync(ideal);
333 }
334 }
335
336 // Helper that guards and inserts a pre-barrier.
337 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
338 Node* pre_val, bool need_mem_bar) const {
339 // We could be accessing the referent field of a reference object. If so, when Shenandoah
340 // is enabled, we need to log the value in the referent field in an SATB buffer.
341 // This routine performs some compile time filters and generates suitable
342 // runtime filters that guard the pre-barrier code.
343 // Also add memory barrier for non volatile load from the referent field
344 // to prevent commoning of loads across safepoint.
345
346 // Some compile time checks.
347
348 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
349 const TypeX* otype = offset->find_intptr_t_type();
350 if (otype != nullptr && otype->is_con() &&
351 otype->get_con() != java_lang_ref_Reference::referent_offset()) {
352 // Constant offset but not the reference_offset so just return
353 return;
354 }
355
356 // We only need to generate the runtime guards for instances.
357 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
358 if (btype != nullptr) {
359 if (btype->isa_aryptr()) {
360 // Array type so nothing to do
361 return;
362 }
363
364 const TypeInstPtr* itype = btype->isa_instptr();
365 if (itype != nullptr) {
366 // Can the klass of base_oop be statically determined to be
367 // _not_ a sub-class of Reference and _not_ Object?
368 ciKlass* klass = itype->instance_klass();
369 if (klass->is_loaded() &&
370 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
371 !kit->env()->Object_klass()->is_subtype_of(klass)) {
372 return;
373 }
374 }
375 }
376
377 // The compile time filters did not reject base_oop/offset so
378 // we need to generate the following runtime filters
379 //
380 // if (offset == java_lang_ref_Reference::_reference_offset) {
381 // if (instance_of(base, java.lang.ref.Reference)) {
382 // pre_barrier(_, pre_val, ...);
383 // }
384 // }
385
386 float likely = PROB_LIKELY( 0.999);
387 float unlikely = PROB_UNLIKELY(0.999);
388
389 IdealKit ideal(kit);
390
391 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset());
392
393 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
394 // Update graphKit memory and control from IdealKit.
395 kit->sync_kit(ideal);
396
397 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
398 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
399
400 // Update IdealKit memory and control from graphKit.
401 __ sync_kit(kit);
402
403 Node* one = __ ConI(1);
404 // is_instof == 0 if base_oop == nullptr
405 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
406
407 // Update graphKit from IdeakKit.
408 kit->sync_kit(ideal);
409
410 // Use the pre-barrier to record the value in the referent field
411 satb_write_barrier_pre(kit, false /* do_load */,
412 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */,
413 pre_val /* pre_val */,
414 T_OBJECT);
415 if (need_mem_bar) {
416 // Add memory barrier to prevent commoning reads from this field
417 // across safepoint since GC can change its value.
418 kit->insert_mem_bar(Op_MemBarCPUOrder);
419 }
420 // Update IdealKit from graphKit.
421 __ sync_kit(kit);
422
423 } __ end_if(); // _ref_type != ref_none
424 } __ end_if(); // offset == referent_offset
425
426 // Final sync IdealKit and GraphKit.
427 kit->final_sync(ideal);
428 }
429
430 Node* ShenandoahBarrierSetC2::byte_map_base_node(GraphKit* kit) const {
431 BarrierSet* bs = BarrierSet::barrier_set();
432 ShenandoahBarrierSet* ctbs = barrier_set_cast<ShenandoahBarrierSet>(bs);
433 CardTable::CardValue* card_table_base = ctbs->card_table()->byte_map_base();
434 if (card_table_base != nullptr) {
435 return kit->makecon(TypeRawPtr::make((address)card_table_base));
436 } else {
437 return kit->null();
438 }
439 }
440
441 void ShenandoahBarrierSetC2::post_barrier(GraphKit* kit,
442 Node* ctl,
443 Node* oop_store,
444 Node* obj,
445 Node* adr,
446 uint adr_idx,
447 Node* val,
448 BasicType bt,
449 bool use_precise) const {
450 assert(ShenandoahCardBarrier, "Should have been checked by caller");
451
452 // No store check needed if we're storing a null.
453 if (val != nullptr && val->is_Con()) {
454 // must be either an oop or NULL
455 const Type* t = val->bottom_type();
456 if (t == TypePtr::NULL_PTR || t == Type::TOP)
457 return;
458 }
459
460 if (ReduceInitialCardMarks && obj == kit->just_allocated_object(kit->control())) {
461 // We can skip marks on a freshly-allocated object in Eden.
462 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
463 // That routine informs GC to take appropriate compensating steps,
464 // upon a slow-path allocation, so as to make this card-mark
465 // elision safe.
466 return;
467 }
468
469 if (!use_precise) {
470 // All card marks for a (non-array) instance are in one place:
471 adr = obj;
472 }
473 // (Else it's an array (or unknown), and we want more precise card marks.)
474 assert(adr != nullptr, "");
475
476 IdealKit ideal(kit, true);
477
478 // Convert the pointer to an int prior to doing math on it
479 Node* cast = __ CastPX(__ ctrl(), adr);
480
481 // Divide by card size
482 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift()) );
483
484 // Combine card table base and card offset
485 Node* card_adr = __ AddP(__ top(), byte_map_base_node(kit), card_offset );
486
487 // Get the alias_index for raw card-mark memory
488 int adr_type = Compile::AliasIdxRaw;
489 Node* zero = __ ConI(0); // Dirty card value
490
491 if (UseCondCardMark) {
492 // The classic GC reference write barrier is typically implemented
493 // as a store into the global card mark table. Unfortunately
494 // unconditional stores can result in false sharing and excessive
495 // coherence traffic as well as false transactional aborts.
496 // UseCondCardMark enables MP "polite" conditional card mark
497 // stores. In theory we could relax the load from ctrl() to
498 // no_ctrl, but that doesn't buy much latitude.
499 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, T_BYTE, adr_type);
500 __ if_then(card_val, BoolTest::ne, zero);
501 }
502
503 // Smash zero into card
504 __ store(__ ctrl(), card_adr, zero, T_BYTE, adr_type, MemNode::unordered);
505
506 if (UseCondCardMark) {
507 __ end_if();
508 }
509
510 // Final sync IdealKit and GraphKit.
511 kit->final_sync(ideal);
512 }
513
514 #undef __
515
516 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() {
517 const Type **fields = TypeTuple::fields(2);
518 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
519 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
520 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
521
522 // create result type (range)
523 fields = TypeTuple::fields(0);
524 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
525
526 return TypeFunc::make(domain, range);
527 }
528
529 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() {
530 const Type **fields = TypeTuple::fields(1);
531 fields[TypeFunc::Parms+0] = TypeOopPtr::NOTNULL; // src oop
532 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
533
534 // create result type (range)
535 fields = TypeTuple::fields(0);
536 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
537
538 return TypeFunc::make(domain, range);
539 }
540
541 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type() {
542 const Type **fields = TypeTuple::fields(2);
543 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM; // original field value
544 fields[TypeFunc::Parms+1] = TypeRawPtr::BOTTOM; // original load address
545
546 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
547
548 // create result type (range)
549 fields = TypeTuple::fields(1);
550 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM;
551 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
552
553 return TypeFunc::make(domain, range);
554 }
555
556 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
557 DecoratorSet decorators = access.decorators();
558
559 const TypePtr* adr_type = access.addr().type();
560 Node* adr = access.addr().node();
561
562 if (!access.is_oop()) {
563 return BarrierSetC2::store_at_resolved(access, val);
564 }
565
566 if (access.is_parse_access()) {
567 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
568 GraphKit* kit = parse_access.kit();
569
570 uint adr_idx = kit->C->get_alias_index(adr_type);
571 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
572 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
573 static_cast<const TypeOopPtr*>(val.type()), nullptr /* pre_val */, access.type());
574
575 Node* result = BarrierSetC2::store_at_resolved(access, val);
576
577 if (ShenandoahCardBarrier) {
578 const bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
579 const bool is_array = (decorators & IS_ARRAY) != 0;
580 const bool use_precise = is_array || anonymous;
581 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
582 adr, adr_idx, val.node(), access.type(), use_precise);
583 }
584 return result;
585 } else {
586 assert(access.is_opt_access(), "only for optimization passes");
587 assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code");
588 return BarrierSetC2::store_at_resolved(access, val);
589 }
590 }
591
592 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
593 // 1: non-reference load, no additional barrier is needed
594 if (!access.is_oop()) {
595 return BarrierSetC2::load_at_resolved(access, val_type);
596 }
597
598 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
599 DecoratorSet decorators = access.decorators();
600 BasicType type = access.type();
601
602 // 2: apply LRB if needed
603 if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
604 load = new ShenandoahLoadReferenceBarrierNode(nullptr, load, decorators);
605 if (access.is_parse_access()) {
606 load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load);
607 } else {
608 load = static_cast<C2OptAccess &>(access).gvn().transform(load);
609 }
610 }
611
612 // 3: apply keep-alive barrier for java.lang.ref.Reference if needed
613 if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
614 Node* top = Compile::current()->top();
615 Node* adr = access.addr().node();
616 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
617 Node* obj = access.base();
618
619 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
620 bool on_weak_ref = (decorators & (ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF)) != 0;
621 bool keep_alive = (decorators & AS_NO_KEEPALIVE) == 0;
622
623 // If we are reading the value of the referent field of a Reference
624 // object (either by using Unsafe directly or through reflection)
625 // then, if SATB is enabled, we need to record the referent in an
626 // SATB log buffer using the pre-barrier mechanism.
627 // Also we need to add memory barrier to prevent commoning reads
628 // from this field across safepoint since GC can change its value.
629 if (!on_weak_ref || (unknown && (offset == top || obj == top)) || !keep_alive) {
630 return load;
631 }
632
633 assert(access.is_parse_access(), "entry not supported at optimization time");
634 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
635 GraphKit* kit = parse_access.kit();
636 bool mismatched = (decorators & C2_MISMATCHED) != 0;
637 bool is_unordered = (decorators & MO_UNORDERED) != 0;
638 bool in_native = (decorators & IN_NATIVE) != 0;
639 bool need_cpu_mem_bar = !is_unordered || mismatched || in_native;
640
641 if (on_weak_ref) {
642 // Use the pre-barrier to record the value in the referent field
643 satb_write_barrier_pre(kit, false /* do_load */,
644 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */,
645 load /* pre_val */, T_OBJECT);
646 // Add memory barrier to prevent commoning reads from this field
647 // across safepoint since GC can change its value.
648 kit->insert_mem_bar(Op_MemBarCPUOrder);
649 } else if (unknown) {
650 // We do not require a mem bar inside pre_barrier if need_mem_bar
651 // is set: the barriers would be emitted by us.
652 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
653 }
654 }
655
656 return load;
657 }
658
659 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
660 Node* new_val, const Type* value_type) const {
661 GraphKit* kit = access.kit();
662 if (access.is_oop()) {
663 shenandoah_write_barrier_pre(kit, false /* do_load */,
664 nullptr, nullptr, max_juint, nullptr, nullptr,
665 expected_val /* pre_val */, T_OBJECT);
666
667 MemNode::MemOrd mo = access.mem_node_mo();
668 Node* mem = access.memory();
669 Node* adr = access.addr().node();
670 const TypePtr* adr_type = access.addr().type();
671 Node* load_store = nullptr;
672
673 #ifdef _LP64
674 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
675 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
676 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
677 if (ShenandoahCASBarrier) {
678 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
679 } else {
680 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
681 }
682 } else
683 #endif
684 {
685 if (ShenandoahCASBarrier) {
686 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
687 } else {
688 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
689 }
690 }
691
692 access.set_raw_access(load_store);
693 pin_atomic_op(access);
694
695 #ifdef _LP64
696 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
697 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
698 }
699 #endif
700 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, load_store, access.decorators()));
701 if (ShenandoahCardBarrier) {
702 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
703 access.addr().node(), access.alias_idx(), new_val, T_OBJECT, true);
704 }
705 return load_store;
706 }
707 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
708 }
709
710 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
711 Node* new_val, const Type* value_type) const {
712 GraphKit* kit = access.kit();
713 if (access.is_oop()) {
714 shenandoah_write_barrier_pre(kit, false /* do_load */,
715 nullptr, nullptr, max_juint, nullptr, nullptr,
716 expected_val /* pre_val */, T_OBJECT);
717 DecoratorSet decorators = access.decorators();
718 MemNode::MemOrd mo = access.mem_node_mo();
719 Node* mem = access.memory();
720 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
721 Node* load_store = nullptr;
722 Node* adr = access.addr().node();
723 #ifdef _LP64
724 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
725 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
726 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
727 if (ShenandoahCASBarrier) {
728 if (is_weak_cas) {
729 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
730 } else {
731 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
732 }
733 } else {
734 if (is_weak_cas) {
735 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
736 } else {
737 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
738 }
739 }
740 } else
741 #endif
742 {
743 if (ShenandoahCASBarrier) {
744 if (is_weak_cas) {
745 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
746 } else {
747 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
748 }
749 } else {
750 if (is_weak_cas) {
751 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
752 } else {
753 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
754 }
755 }
756 }
757 access.set_raw_access(load_store);
758 pin_atomic_op(access);
759 if (ShenandoahCardBarrier) {
760 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
761 access.addr().node(), access.alias_idx(), new_val, T_OBJECT, true);
762 }
763 return load_store;
764 }
765 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
766 }
767
768 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
769 GraphKit* kit = access.kit();
770 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
771 if (access.is_oop()) {
772 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, result, access.decorators()));
773 shenandoah_write_barrier_pre(kit, false /* do_load */,
774 nullptr, nullptr, max_juint, nullptr, nullptr,
775 result /* pre_val */, T_OBJECT);
776 if (ShenandoahCardBarrier) {
777 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
778 access.addr().node(), access.alias_idx(), val, T_OBJECT, true);
779 }
780 }
781 return result;
782 }
783
784
785 bool ShenandoahBarrierSetC2::is_gc_pre_barrier_node(Node* node) const {
786 return is_shenandoah_wb_pre_call(node);
787 }
788
789 // Support for GC barriers emitted during parsing
790 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
791 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) return true;
792 if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) {
793 return false;
794 }
795 CallLeafNode *call = node->as_CallLeaf();
796 if (call->_name == nullptr) {
797 return false;
798 }
799
800 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 ||
801 strcmp(call->_name, "shenandoah_cas_obj") == 0 ||
802 strcmp(call->_name, "shenandoah_wb_pre") == 0;
803 }
804
805 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
806 if (c == nullptr) {
807 return c;
808 }
809 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
810 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
811 }
812 return c;
813 }
814
815 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
816 return !ShenandoahBarrierC2Support::expand(C, igvn);
817 }
818
819 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
820 if (mode == LoopOptsShenandoahExpand) {
821 assert(UseShenandoahGC, "only for shenandoah");
822 ShenandoahBarrierC2Support::pin_and_expand(phase);
823 return true;
824 } else if (mode == LoopOptsShenandoahPostExpand) {
825 assert(UseShenandoahGC, "only for shenandoah");
826 visited.clear();
827 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
828 return true;
829 }
830 return false;
831 }
832
833 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, bool is_clone_instance, ArrayCopyPhase phase) const {
834 bool is_oop = is_reference_type(type);
835 if (!is_oop) {
836 return false;
837 }
838 if (ShenandoahSATBBarrier && tightly_coupled_alloc) {
839 if (phase == Optimization) {
840 return false;
841 }
842 return !is_clone;
843 }
844 return true;
845 }
846
847 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
848 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
849 if (src_type->isa_instptr() != nullptr) {
850 ciInstanceKlass* ik = src_type->is_instptr()->instance_klass();
851 if ((src_type->klass_is_exact() || !ik->has_subklass()) && !ik->has_injected_fields()) {
852 if (ik->has_object_fields()) {
853 return true;
854 } else {
855 if (!src_type->klass_is_exact()) {
856 Compile::current()->dependencies()->assert_leaf_type(ik);
857 }
858 }
859 } else {
860 return true;
861 }
862 } else if (src_type->isa_aryptr()) {
863 BasicType src_elem = src_type->isa_aryptr()->elem()->array_element_basic_type();
864 if (is_reference_type(src_elem, true)) {
865 return true;
866 }
867 } else {
868 return true;
869 }
870 return false;
871 }
872
873 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
874 Node* ctrl = ac->in(TypeFunc::Control);
875 Node* mem = ac->in(TypeFunc::Memory);
876 Node* src_base = ac->in(ArrayCopyNode::Src);
877 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
878 Node* dest_base = ac->in(ArrayCopyNode::Dest);
879 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
880 Node* length = ac->in(ArrayCopyNode::Length);
881
882 Node* src = phase->basic_plus_adr(src_base, src_offset);
883 Node* dest = phase->basic_plus_adr(dest_base, dest_offset);
884
885 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
886 // Check if heap is has forwarded objects. If it does, we need to call into the special
887 // routine that would fix up source references before we can continue.
888
889 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
890 Node* region = new RegionNode(PATH_LIMIT);
891 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM);
892
893 Node* thread = phase->transform_later(new ThreadLocalNode());
894 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
895 Node* gc_state_addr = phase->transform_later(new AddPNode(phase->C->top(), thread, offset));
896
897 uint gc_state_idx = Compile::AliasIdxRaw;
898 const TypePtr* gc_state_adr_type = nullptr; // debug-mode-only argument
899 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
900
901 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered));
902 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED)));
903 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT)));
904 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne));
905
906 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If();
907 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff));
908 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff));
909
910 // Heap is stable, no need to do anything additional
911 region->init_req(_heap_stable, stable_ctrl);
912 mem_phi->init_req(_heap_stable, mem);
913
914 // Heap is unstable, call into clone barrier stub
915 Node* call = phase->make_leaf_call(unstable_ctrl, mem,
916 ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(),
917 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier),
918 "shenandoah_clone",
919 TypeRawPtr::BOTTOM,
920 src_base);
921 call = phase->transform_later(call);
922
923 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control));
924 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory));
925 region->init_req(_heap_unstable, ctrl);
926 mem_phi->init_req(_heap_unstable, mem);
927
928 // Wire up the actual arraycopy stub now
929 ctrl = phase->transform_later(region);
930 mem = phase->transform_later(mem_phi);
931
932 const char* name = "arraycopy";
933 call = phase->make_leaf_call(ctrl, mem,
934 OptoRuntime::fast_arraycopy_Type(),
935 phase->basictype2arraycopy(T_LONG, nullptr, nullptr, true, name, true),
936 name, TypeRawPtr::BOTTOM,
937 src, dest, length
938 LP64_ONLY(COMMA phase->top()));
939 call = phase->transform_later(call);
940
941 // Hook up the whole thing into the graph
942 phase->igvn().replace_node(ac, call);
943 } else {
944 BarrierSetC2::clone_at_expansion(phase, ac);
945 }
946 }
947
948
949 // Support for macro expanded GC barriers
950 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
951 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
952 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
953 }
954 }
955
956 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
957 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
958 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
959 }
960 }
961
962 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const {
963 if (is_shenandoah_wb_pre_call(node)) {
964 shenandoah_eliminate_wb_pre(node, ¯o->igvn());
965 }
966 if (ShenandoahCardBarrier && node->Opcode() == Op_CastP2X) {
967 Node* shift = node->unique_out();
968 Node* addp = shift->unique_out();
969 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
970 Node* mem = addp->last_out(j);
971 if (UseCondCardMark && mem->is_Load()) {
972 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
973 // The load is checking if the card has been written so
974 // replace it with zero to fold the test.
975 macro->replace_node(mem, macro->intcon(0));
976 continue;
977 }
978 assert(mem->is_Store(), "store required");
979 macro->replace_node(mem, mem->in(MemNode::Memory));
980 }
981 }
982 }
983
984 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
985 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
986 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
987 c = c->unique_ctrl_out();
988 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
989 c = c->unique_ctrl_out();
990 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
991 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
992 assert(iff->is_If(), "expect test");
993 if (!is_shenandoah_marking_if(igvn, iff)) {
994 c = c->unique_ctrl_out();
995 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
996 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
997 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
998 }
999 Node* cmpx = iff->in(1)->in(1);
1000 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
1001 igvn->rehash_node_delayed(call);
1002 call->del_req(call->req()-1);
1003 }
1004
1005 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
1006 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
1007 igvn->add_users_to_worklist(node);
1008 }
1009 }
1010
1011 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
1012 for (uint i = 0; i < useful.size(); i++) {
1013 Node* n = useful.at(i);
1014 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
1015 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1016 C->record_for_igvn(n->fast_out(i));
1017 }
1018 }
1019 }
1020
1021 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
1022 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
1023 if (!useful.member(n)) {
1024 state()->remove_load_reference_barrier(n);
1025 }
1026 }
1027 }
1028
1029 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
1030 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
1031 }
1032
1033 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
1034 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
1035 }
1036
1037 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
1038 // expanded later, then now is the time to do so.
1039 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
1040
1041 #ifdef ASSERT
1042 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
1043 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) {
1044 ShenandoahBarrierC2Support::verify(Compile::current()->root());
1045 } else if (phase == BarrierSetC2::BeforeCodeGen) {
1046 // Verify Shenandoah pre-barriers
1047 const int gc_state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
1048
1049 Unique_Node_List visited;
1050 Node_List worklist;
1051 // We're going to walk control flow backwards starting from the Root
1052 worklist.push(compile->root());
1053 while (worklist.size() > 0) {
1054 Node *x = worklist.pop();
1055 if (x == nullptr || x == compile->top()) {
1056 continue;
1057 }
1058
1059 if (visited.member(x)) {
1060 continue;
1061 } else {
1062 visited.push(x);
1063 }
1064
1065 if (x->is_Region()) {
1066 for (uint i = 1; i < x->req(); i++) {
1067 worklist.push(x->in(i));
1068 }
1069 } else {
1070 worklist.push(x->in(0));
1071 // We are looking for the pattern:
1072 // /->ThreadLocal
1073 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
1074 // \->ConI(0)
1075 // We want to verify that the If and the LoadB have the same control
1076 // See GraphKit::g1_write_barrier_pre()
1077 if (x->is_If()) {
1078 IfNode *iff = x->as_If();
1079 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
1080 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
1081 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
1082 && cmp->in(1)->is_Load()) {
1083 LoadNode *load = cmp->in(1)->as_Load();
1084 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
1085 && load->in(2)->in(3)->is_Con()
1086 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == gc_state_offset) {
1087
1088 Node *if_ctrl = iff->in(0);
1089 Node *load_ctrl = load->in(0);
1090
1091 if (if_ctrl != load_ctrl) {
1092 // Skip possible CProj->NeverBranch in infinite loops
1093 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
1094 && if_ctrl->in(0)->is_NeverBranch()) {
1095 if_ctrl = if_ctrl->in(0)->in(0);
1096 }
1097 }
1098 assert(load_ctrl != nullptr && if_ctrl == load_ctrl, "controls must match");
1099 }
1100 }
1101 }
1102 }
1103 }
1104 }
1105 }
1106 }
1107 #endif
1108
1109 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
1110 if (is_shenandoah_wb_pre_call(n)) {
1111 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1112 if (n->req() > cnt) {
1113 Node* addp = n->in(cnt);
1114 if (has_only_shenandoah_wb_pre_uses(addp)) {
1115 n->del_req(cnt);
1116 if (can_reshape) {
1117 phase->is_IterGVN()->_worklist.push(addp);
1118 }
1119 return n;
1120 }
1121 }
1122 }
1123 if (n->Opcode() == Op_CmpP) {
1124 Node* in1 = n->in(1);
1125 Node* in2 = n->in(2);
1126
1127 // If one input is null, then step over the strong LRB barriers on the other input
1128 if (in1->bottom_type() == TypePtr::NULL_PTR &&
1129 !((in2->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1130 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in2)->decorators()))) {
1131 in2 = step_over_gc_barrier(in2);
1132 }
1133 if (in2->bottom_type() == TypePtr::NULL_PTR &&
1134 !((in1->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1135 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in1)->decorators()))) {
1136 in1 = step_over_gc_barrier(in1);
1137 }
1138
1139 if (in1 != n->in(1)) {
1140 n->set_req_X(1, in1, phase);
1141 assert(in2 == n->in(2), "only one change");
1142 return n;
1143 }
1144 if (in2 != n->in(2)) {
1145 n->set_req_X(2, in2, phase);
1146 return n;
1147 }
1148 } else if (can_reshape &&
1149 n->Opcode() == Op_If &&
1150 ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1151 n->in(0) != nullptr &&
1152 n->outcnt() == 2) {
1153 Node* dom = n->in(0);
1154 Node* prev_dom = n;
1155 int op = n->Opcode();
1156 int dist = 16;
1157 // Search up the dominator tree for another heap stable test
1158 while (dom->Opcode() != op || // Not same opcode?
1159 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1?
1160 prev_dom->in(0) != dom) { // One path of test does not dominate?
1161 if (dist < 0) return nullptr;
1162
1163 dist--;
1164 prev_dom = dom;
1165 dom = IfNode::up_one_dom(dom);
1166 if (!dom) return nullptr;
1167 }
1168
1169 // Check that we did not follow a loop back to ourselves
1170 if (n == dom) {
1171 return nullptr;
1172 }
1173
1174 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN(), false);
1175 }
1176
1177 return nullptr;
1178 }
1179
1180 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1181 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1182 Node* u = n->fast_out(i);
1183 if (!is_shenandoah_wb_pre_call(u)) {
1184 return false;
1185 }
1186 }
1187 return n->outcnt() > 0;
1188 }
1189
1190 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode, Unique_Node_List& dead_nodes) const {
1191 switch (opcode) {
1192 case Op_CallLeaf:
1193 case Op_CallLeafNoFP: {
1194 assert (n->is_Call(), "");
1195 CallNode *call = n->as_Call();
1196 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1197 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1198 if (call->req() > cnt) {
1199 assert(call->req() == cnt + 1, "only one extra input");
1200 Node *addp = call->in(cnt);
1201 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1202 call->del_req(cnt);
1203 }
1204 }
1205 return false;
1206 }
1207 case Op_ShenandoahCompareAndSwapP:
1208 case Op_ShenandoahCompareAndSwapN:
1209 case Op_ShenandoahWeakCompareAndSwapN:
1210 case Op_ShenandoahWeakCompareAndSwapP:
1211 case Op_ShenandoahCompareAndExchangeP:
1212 case Op_ShenandoahCompareAndExchangeN:
1213 return true;
1214 case Op_ShenandoahLoadReferenceBarrier:
1215 assert(false, "should have been expanded already");
1216 return true;
1217 default:
1218 return false;
1219 }
1220 }
1221
1222 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1223 switch (opcode) {
1224 case Op_ShenandoahCompareAndExchangeP:
1225 case Op_ShenandoahCompareAndExchangeN:
1226 conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1227 // fallthrough
1228 case Op_ShenandoahWeakCompareAndSwapP:
1229 case Op_ShenandoahWeakCompareAndSwapN:
1230 case Op_ShenandoahCompareAndSwapP:
1231 case Op_ShenandoahCompareAndSwapN:
1232 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1233 return true;
1234 case Op_StoreP: {
1235 Node* adr = n->in(MemNode::Address);
1236 const Type* adr_type = gvn->type(adr);
1237 // Pointer stores in Shenandoah barriers looks like unsafe access.
1238 // Ignore such stores to be able scalar replace non-escaping
1239 // allocations.
1240 if (adr_type->isa_rawptr() && adr->is_AddP()) {
1241 Node* base = conn_graph->get_addp_base(adr);
1242 if (base->Opcode() == Op_LoadP &&
1243 base->in(MemNode::Address)->is_AddP()) {
1244 adr = base->in(MemNode::Address);
1245 Node* tls = conn_graph->get_addp_base(adr);
1246 if (tls->Opcode() == Op_ThreadLocal) {
1247 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1248 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1249 if (offs == buf_offset) {
1250 return true; // Pre barrier previous oop value store.
1251 }
1252 }
1253 }
1254 }
1255 return false;
1256 }
1257 case Op_ShenandoahLoadReferenceBarrier:
1258 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist);
1259 return true;
1260 default:
1261 // Nothing
1262 break;
1263 }
1264 return false;
1265 }
1266
1267 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1268 switch (opcode) {
1269 case Op_ShenandoahCompareAndExchangeP:
1270 case Op_ShenandoahCompareAndExchangeN: {
1271 Node *adr = n->in(MemNode::Address);
1272 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, nullptr);
1273 // fallthrough
1274 }
1275 case Op_ShenandoahCompareAndSwapP:
1276 case Op_ShenandoahCompareAndSwapN:
1277 case Op_ShenandoahWeakCompareAndSwapP:
1278 case Op_ShenandoahWeakCompareAndSwapN:
1279 return conn_graph->add_final_edges_unsafe_access(n, opcode);
1280 case Op_ShenandoahLoadReferenceBarrier:
1281 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), nullptr);
1282 return true;
1283 default:
1284 // Nothing
1285 break;
1286 }
1287 return false;
1288 }
1289
1290 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1291 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1292 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1293
1294 }
1295
1296 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1297 switch (opcode) {
1298 case Op_ShenandoahCompareAndExchangeP:
1299 case Op_ShenandoahCompareAndExchangeN:
1300 case Op_ShenandoahWeakCompareAndSwapP:
1301 case Op_ShenandoahWeakCompareAndSwapN:
1302 case Op_ShenandoahCompareAndSwapP:
1303 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree
1304 Node* newval = n->in(MemNode::ValueIn);
1305 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1306 Node* pair = new BinaryNode(oldval, newval);
1307 n->set_req(MemNode::ValueIn,pair);
1308 n->del_req(LoadStoreConditionalNode::ExpectedIn);
1309 return true;
1310 }
1311 default:
1312 break;
1313 }
1314 return false;
1315 }
1316
1317 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1318 return xop == Op_ShenandoahCompareAndExchangeP ||
1319 xop == Op_ShenandoahCompareAndExchangeN ||
1320 xop == Op_ShenandoahWeakCompareAndSwapP ||
1321 xop == Op_ShenandoahWeakCompareAndSwapN ||
1322 xop == Op_ShenandoahCompareAndSwapN ||
1323 xop == Op_ShenandoahCompareAndSwapP;
1324 }