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, phase);
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, phase),
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_clone_call(Node* call) {
277 return call->is_CallLeaf() &&
278 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier);
279 }
280
281 bool ShenandoahBarrierSetC2::is_shenandoah_lrb_call(Node* call) {
282 if (!call->is_CallLeaf()) {
283 return false;
284 }
285
286 address entry_point = call->as_CallLeaf()->entry_point();
287 return (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong)) ||
288 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow)) ||
289 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak)) ||
290 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow)) ||
291 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom)) ||
292 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow));
293 }
294
295 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseValues* phase, Node* n) {
296 if (n->Opcode() != Op_If) {
297 return false;
298 }
299
300 Node* bol = n->in(1);
301 assert(bol->is_Bool(), "");
302 Node* cmpx = bol->in(1);
303 if (bol->as_Bool()->_test._test == BoolTest::ne &&
304 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
305 is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
306 cmpx->in(1)->in(2)->is_Con() &&
307 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
308 return true;
309 }
310
311 return false;
312 }
313
314 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
315 if (!n->is_Load()) return false;
316 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
317 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
318 && n->in(2)->in(3)->is_Con()
319 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
320 }
321
322 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
323 bool do_load,
324 Node* obj,
325 Node* adr,
326 uint alias_idx,
327 Node* val,
328 const TypeOopPtr* val_type,
329 Node* pre_val,
330 BasicType bt) const {
331 if (ShenandoahSATBBarrier) {
332 IdealKit ideal(kit);
333 kit->sync_kit(ideal);
334
335 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
336
337 ideal.sync_kit(kit);
338 kit->final_sync(ideal);
339 }
340 }
341
342 // Helper that guards and inserts a pre-barrier.
343 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
344 Node* pre_val, bool need_mem_bar) const {
345 // We could be accessing the referent field of a reference object. If so, when Shenandoah
346 // is enabled, we need to log the value in the referent field in an SATB buffer.
347 // This routine performs some compile time filters and generates suitable
348 // runtime filters that guard the pre-barrier code.
349 // Also add memory barrier for non volatile load from the referent field
350 // to prevent commoning of loads across safepoint.
351
352 // Some compile time checks.
353
354 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
355 const TypeX* otype = offset->find_intptr_t_type();
356 if (otype != nullptr && otype->is_con() &&
357 otype->get_con() != java_lang_ref_Reference::referent_offset()) {
358 // Constant offset but not the reference_offset so just return
359 return;
360 }
361
362 // We only need to generate the runtime guards for instances.
363 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
364 if (btype != nullptr) {
365 if (btype->isa_aryptr()) {
366 // Array type so nothing to do
367 return;
368 }
369
370 const TypeInstPtr* itype = btype->isa_instptr();
371 if (itype != nullptr) {
372 // Can the klass of base_oop be statically determined to be
373 // _not_ a sub-class of Reference and _not_ Object?
374 ciKlass* klass = itype->instance_klass();
375 if (klass->is_loaded() &&
376 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
377 !kit->env()->Object_klass()->is_subtype_of(klass)) {
378 return;
379 }
380 }
381 }
382
383 // The compile time filters did not reject base_oop/offset so
384 // we need to generate the following runtime filters
385 //
386 // if (offset == java_lang_ref_Reference::_reference_offset) {
387 // if (instance_of(base, java.lang.ref.Reference)) {
388 // pre_barrier(_, pre_val, ...);
389 // }
390 // }
391
392 float likely = PROB_LIKELY( 0.999);
393 float unlikely = PROB_UNLIKELY(0.999);
394
395 IdealKit ideal(kit);
396
397 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset());
398
399 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
400 // Update graphKit memory and control from IdealKit.
401 kit->sync_kit(ideal);
402
403 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
404 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
405
406 // Update IdealKit memory and control from graphKit.
407 __ sync_kit(kit);
408
409 Node* one = __ ConI(1);
410 // is_instof == 0 if base_oop == nullptr
411 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
412
413 // Update graphKit from IdeakKit.
414 kit->sync_kit(ideal);
415
416 // Use the pre-barrier to record the value in the referent field
417 satb_write_barrier_pre(kit, false /* do_load */,
418 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */,
419 pre_val /* pre_val */,
420 T_OBJECT);
421 if (need_mem_bar) {
422 // Add memory barrier to prevent commoning reads from this field
423 // across safepoint since GC can change its value.
424 kit->insert_mem_bar(Op_MemBarCPUOrder);
425 }
426 // Update IdealKit from graphKit.
427 __ sync_kit(kit);
428
429 } __ end_if(); // _ref_type != ref_none
430 } __ end_if(); // offset == referent_offset
431
432 // Final sync IdealKit and GraphKit.
433 kit->final_sync(ideal);
434 }
435
436 Node* ShenandoahBarrierSetC2::byte_map_base_node(GraphKit* kit) const {
437 BarrierSet* bs = BarrierSet::barrier_set();
438 ShenandoahBarrierSet* ctbs = barrier_set_cast<ShenandoahBarrierSet>(bs);
439 CardTable::CardValue* card_table_base = ctbs->card_table()->byte_map_base();
440 if (card_table_base != nullptr) {
441 return kit->makecon(TypeRawPtr::make((address)card_table_base));
442 } else {
443 return kit->null();
444 }
445 }
446
447 void ShenandoahBarrierSetC2::post_barrier(GraphKit* kit,
448 Node* ctl,
449 Node* oop_store,
450 Node* obj,
451 Node* adr,
452 uint adr_idx,
453 Node* val,
454 BasicType bt,
455 bool use_precise) const {
456 assert(ShenandoahCardBarrier, "Should have been checked by caller");
457
458 // No store check needed if we're storing a null.
459 if (val != nullptr && val->is_Con()) {
460 // must be either an oop or NULL
461 const Type* t = val->bottom_type();
462 if (t == TypePtr::NULL_PTR || t == Type::TOP)
463 return;
464 }
465
466 if (ReduceInitialCardMarks && obj == kit->just_allocated_object(kit->control())) {
467 // We can skip marks on a freshly-allocated object in Eden.
468 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
469 // That routine informs GC to take appropriate compensating steps,
470 // upon a slow-path allocation, so as to make this card-mark
471 // elision safe.
472 return;
473 }
474
475 if (!use_precise) {
476 // All card marks for a (non-array) instance are in one place:
477 adr = obj;
478 }
479 // (Else it's an array (or unknown), and we want more precise card marks.)
480 assert(adr != nullptr, "");
481
482 IdealKit ideal(kit, true);
483
484 // Convert the pointer to an int prior to doing math on it
485 Node* cast = __ CastPX(__ ctrl(), adr);
486
487 // Divide by card size
488 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift()) );
489
490 // Combine card table base and card offset
491 Node* card_adr = __ AddP(__ top(), byte_map_base_node(kit), card_offset );
492
493 // Get the alias_index for raw card-mark memory
494 int adr_type = Compile::AliasIdxRaw;
495 Node* zero = __ ConI(0); // Dirty card value
496
497 if (UseCondCardMark) {
498 // The classic GC reference write barrier is typically implemented
499 // as a store into the global card mark table. Unfortunately
500 // unconditional stores can result in false sharing and excessive
501 // coherence traffic as well as false transactional aborts.
502 // UseCondCardMark enables MP "polite" conditional card mark
503 // stores. In theory we could relax the load from ctrl() to
504 // no_ctrl, but that doesn't buy much latitude.
505 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, T_BYTE, adr_type);
506 __ if_then(card_val, BoolTest::ne, zero);
507 }
508
509 // Smash zero into card
510 __ store(__ ctrl(), card_adr, zero, T_BYTE, adr_type, MemNode::unordered);
511
512 if (UseCondCardMark) {
513 __ end_if();
514 }
515
516 // Final sync IdealKit and GraphKit.
517 kit->final_sync(ideal);
518 }
519
520 #undef __
521
522 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() {
523 const Type **fields = TypeTuple::fields(2);
524 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
525 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
526 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
527
528 // create result type (range)
529 fields = TypeTuple::fields(0);
530 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
531
532 return TypeFunc::make(domain, range);
533 }
534
535 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() {
536 const Type **fields = TypeTuple::fields(1);
537 fields[TypeFunc::Parms+0] = TypeOopPtr::NOTNULL; // src oop
538 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
539
540 // create result type (range)
541 fields = TypeTuple::fields(0);
542 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
543
544 return TypeFunc::make(domain, range);
545 }
546
547 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type() {
548 const Type **fields = TypeTuple::fields(2);
549 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM; // original field value
550 fields[TypeFunc::Parms+1] = TypeRawPtr::BOTTOM; // original load address
551
552 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
553
554 // create result type (range)
555 fields = TypeTuple::fields(1);
556 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM;
557 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
558
559 return TypeFunc::make(domain, range);
560 }
561
562 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
563 DecoratorSet decorators = access.decorators();
564
565 const TypePtr* adr_type = access.addr().type();
566 Node* adr = access.addr().node();
567
568 if (!access.is_oop()) {
569 return BarrierSetC2::store_at_resolved(access, val);
570 }
571
572 if (access.is_parse_access()) {
573 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
574 GraphKit* kit = parse_access.kit();
575
576 uint adr_idx = kit->C->get_alias_index(adr_type);
577 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
578 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
579 static_cast<const TypeOopPtr*>(val.type()), nullptr /* pre_val */, access.type());
580
581 Node* result = BarrierSetC2::store_at_resolved(access, val);
582
583 if (ShenandoahCardBarrier) {
584 const bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
585 const bool is_array = (decorators & IS_ARRAY) != 0;
586 const bool use_precise = is_array || anonymous;
587 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
588 adr, adr_idx, val.node(), access.type(), use_precise);
589 }
590 return result;
591 } else {
592 assert(access.is_opt_access(), "only for optimization passes");
593 assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code");
594 return BarrierSetC2::store_at_resolved(access, val);
595 }
596 }
597
598 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
599 // 1: non-reference load, no additional barrier is needed
600 if (!access.is_oop()) {
601 return BarrierSetC2::load_at_resolved(access, val_type);
602 }
603
604 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
605 DecoratorSet decorators = access.decorators();
606 BasicType type = access.type();
607
608 // 2: apply LRB if needed
609 if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
610 load = new ShenandoahLoadReferenceBarrierNode(nullptr, load, decorators);
611 if (access.is_parse_access()) {
612 load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load);
613 } else {
614 load = static_cast<C2OptAccess &>(access).gvn().transform(load);
615 }
616 }
617
618 // 3: apply keep-alive barrier for java.lang.ref.Reference if needed
619 if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
620 Node* top = Compile::current()->top();
621 Node* adr = access.addr().node();
622 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
623 Node* obj = access.base();
624
625 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
626 bool on_weak_ref = (decorators & (ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF)) != 0;
627 bool keep_alive = (decorators & AS_NO_KEEPALIVE) == 0;
628
629 // If we are reading the value of the referent field of a Reference
630 // object (either by using Unsafe directly or through reflection)
631 // then, if SATB is enabled, we need to record the referent in an
632 // SATB log buffer using the pre-barrier mechanism.
633 // Also we need to add memory barrier to prevent commoning reads
634 // from this field across safepoint since GC can change its value.
635 if (!on_weak_ref || (unknown && (offset == top || obj == top)) || !keep_alive) {
636 return load;
637 }
638
639 assert(access.is_parse_access(), "entry not supported at optimization time");
640 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
641 GraphKit* kit = parse_access.kit();
642 bool mismatched = (decorators & C2_MISMATCHED) != 0;
643 bool is_unordered = (decorators & MO_UNORDERED) != 0;
644 bool in_native = (decorators & IN_NATIVE) != 0;
645 bool need_cpu_mem_bar = !is_unordered || mismatched || in_native;
646
647 if (on_weak_ref) {
648 // Use the pre-barrier to record the value in the referent field
649 satb_write_barrier_pre(kit, false /* do_load */,
650 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */,
651 load /* pre_val */, T_OBJECT);
652 // Add memory barrier to prevent commoning reads from this field
653 // across safepoint since GC can change its value.
654 kit->insert_mem_bar(Op_MemBarCPUOrder);
655 } else if (unknown) {
656 // We do not require a mem bar inside pre_barrier if need_mem_bar
657 // is set: the barriers would be emitted by us.
658 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
659 }
660 }
661
662 return load;
663 }
664
665 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
666 Node* new_val, const Type* value_type) const {
667 GraphKit* kit = access.kit();
668 if (access.is_oop()) {
669 shenandoah_write_barrier_pre(kit, false /* do_load */,
670 nullptr, nullptr, max_juint, nullptr, nullptr,
671 expected_val /* pre_val */, T_OBJECT);
672
673 MemNode::MemOrd mo = access.mem_node_mo();
674 Node* mem = access.memory();
675 Node* adr = access.addr().node();
676 const TypePtr* adr_type = access.addr().type();
677 Node* load_store = nullptr;
678
679 #ifdef _LP64
680 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
681 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
682 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
683 if (ShenandoahCASBarrier) {
684 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
685 } else {
686 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
687 }
688 } else
689 #endif
690 {
691 if (ShenandoahCASBarrier) {
692 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
693 } else {
694 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
695 }
696 }
697
698 access.set_raw_access(load_store);
699 pin_atomic_op(access);
700
701 #ifdef _LP64
702 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
703 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
704 }
705 #endif
706 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, load_store, access.decorators()));
707 if (ShenandoahCardBarrier) {
708 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
709 access.addr().node(), access.alias_idx(), new_val, T_OBJECT, true);
710 }
711 return load_store;
712 }
713 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
714 }
715
716 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
717 Node* new_val, const Type* value_type) const {
718 GraphKit* kit = access.kit();
719 if (access.is_oop()) {
720 shenandoah_write_barrier_pre(kit, false /* do_load */,
721 nullptr, nullptr, max_juint, nullptr, nullptr,
722 expected_val /* pre_val */, T_OBJECT);
723 DecoratorSet decorators = access.decorators();
724 MemNode::MemOrd mo = access.mem_node_mo();
725 Node* mem = access.memory();
726 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
727 Node* load_store = nullptr;
728 Node* adr = access.addr().node();
729 #ifdef _LP64
730 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
731 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
732 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
733 if (ShenandoahCASBarrier) {
734 if (is_weak_cas) {
735 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
736 } else {
737 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
738 }
739 } else {
740 if (is_weak_cas) {
741 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
742 } else {
743 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
744 }
745 }
746 } else
747 #endif
748 {
749 if (ShenandoahCASBarrier) {
750 if (is_weak_cas) {
751 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
752 } else {
753 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
754 }
755 } else {
756 if (is_weak_cas) {
757 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
758 } else {
759 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
760 }
761 }
762 }
763 access.set_raw_access(load_store);
764 pin_atomic_op(access);
765 if (ShenandoahCardBarrier) {
766 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
767 access.addr().node(), access.alias_idx(), new_val, T_OBJECT, true);
768 }
769 return load_store;
770 }
771 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
772 }
773
774 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
775 GraphKit* kit = access.kit();
776 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
777 if (access.is_oop()) {
778 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, result, access.decorators()));
779 shenandoah_write_barrier_pre(kit, false /* do_load */,
780 nullptr, nullptr, max_juint, nullptr, nullptr,
781 result /* pre_val */, T_OBJECT);
782 if (ShenandoahCardBarrier) {
783 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
784 access.addr().node(), access.alias_idx(), val, T_OBJECT, true);
785 }
786 }
787 return result;
788 }
789
790
791 bool ShenandoahBarrierSetC2::is_gc_pre_barrier_node(Node* node) const {
792 return is_shenandoah_wb_pre_call(node);
793 }
794
795 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
796 return (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) ||
797 is_shenandoah_lrb_call(node) ||
798 is_shenandoah_wb_pre_call(node) ||
799 is_shenandoah_clone_call(node);
800 }
801
802 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
803 if (c == nullptr) {
804 return c;
805 }
806 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
807 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
808 }
809 return c;
810 }
811
812 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
813 return !ShenandoahBarrierC2Support::expand(C, igvn);
814 }
815
816 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
817 if (mode == LoopOptsShenandoahExpand) {
818 assert(UseShenandoahGC, "only for shenandoah");
819 ShenandoahBarrierC2Support::pin_and_expand(phase);
820 return true;
821 } else if (mode == LoopOptsShenandoahPostExpand) {
822 assert(UseShenandoahGC, "only for shenandoah");
823 visited.clear();
824 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
825 return true;
826 }
827 return false;
828 }
829
830 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, bool is_clone_instance, ArrayCopyPhase phase) const {
831 bool is_oop = is_reference_type(type);
832 if (!is_oop) {
833 return false;
834 }
835 if (ShenandoahSATBBarrier && tightly_coupled_alloc) {
836 if (phase == Optimization) {
837 return false;
838 }
839 return !is_clone;
840 }
841 return true;
842 }
843
844 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
845 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
846 if (src_type->isa_instptr() != nullptr) {
847 ciInstanceKlass* ik = src_type->is_instptr()->instance_klass();
848 if ((src_type->klass_is_exact() || !ik->has_subklass()) && !ik->has_injected_fields()) {
849 if (ik->has_object_fields()) {
850 return true;
851 } else {
852 if (!src_type->klass_is_exact()) {
853 Compile::current()->dependencies()->assert_leaf_type(ik);
854 }
855 }
856 } else {
857 return true;
858 }
859 } else if (src_type->isa_aryptr()) {
860 BasicType src_elem = src_type->isa_aryptr()->elem()->array_element_basic_type();
861 if (is_reference_type(src_elem, true)) {
862 return true;
863 }
864 } else {
865 return true;
866 }
867 return false;
868 }
869
870 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
871 Node* ctrl = ac->in(TypeFunc::Control);
872 Node* mem = ac->in(TypeFunc::Memory);
873 Node* src_base = ac->in(ArrayCopyNode::Src);
874 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
875 Node* dest_base = ac->in(ArrayCopyNode::Dest);
876 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
877 Node* length = ac->in(ArrayCopyNode::Length);
878
879 Node* src = phase->basic_plus_adr(src_base, src_offset);
880 Node* dest = phase->basic_plus_adr(dest_base, dest_offset);
881
882 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
883 // Check if heap is has forwarded objects. If it does, we need to call into the special
884 // routine that would fix up source references before we can continue.
885
886 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
887 Node* region = new RegionNode(PATH_LIMIT);
888 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM);
889
890 Node* thread = phase->transform_later(new ThreadLocalNode());
891 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
892 Node* gc_state_addr = phase->transform_later(new AddPNode(phase->C->top(), thread, offset));
893
894 uint gc_state_idx = Compile::AliasIdxRaw;
895 const TypePtr* gc_state_adr_type = nullptr; // debug-mode-only argument
896 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
897
898 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered));
899 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED)));
900 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT)));
901 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne));
902
903 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If();
904 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff));
905 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff));
906
907 // Heap is stable, no need to do anything additional
908 region->init_req(_heap_stable, stable_ctrl);
909 mem_phi->init_req(_heap_stable, mem);
910
911 // Heap is unstable, call into clone barrier stub
912 Node* call = phase->make_leaf_call(unstable_ctrl, mem,
913 ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(),
914 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier),
915 "shenandoah_clone",
916 TypeRawPtr::BOTTOM,
917 src_base);
918 call = phase->transform_later(call);
919
920 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control));
921 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory));
922 region->init_req(_heap_unstable, ctrl);
923 mem_phi->init_req(_heap_unstable, mem);
924
925 // Wire up the actual arraycopy stub now
926 ctrl = phase->transform_later(region);
927 mem = phase->transform_later(mem_phi);
928
929 const char* name = "arraycopy";
930 call = phase->make_leaf_call(ctrl, mem,
931 OptoRuntime::fast_arraycopy_Type(),
932 phase->basictype2arraycopy(T_LONG, nullptr, nullptr, true, name, true),
933 name, TypeRawPtr::BOTTOM,
934 src, dest, length
935 LP64_ONLY(COMMA phase->top()));
936 call = phase->transform_later(call);
937
938 // Hook up the whole thing into the graph
939 phase->igvn().replace_node(ac, call);
940 } else {
941 BarrierSetC2::clone_at_expansion(phase, ac);
942 }
943 }
944
945
946 // Support for macro expanded GC barriers
947 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
948 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
949 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
950 }
951 }
952
953 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
954 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
955 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
956 }
957 }
958
959 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const {
960 if (is_shenandoah_wb_pre_call(node)) {
961 shenandoah_eliminate_wb_pre(node, ¯o->igvn());
962 }
963 if (ShenandoahCardBarrier && node->Opcode() == Op_CastP2X) {
964 Node* shift = node->unique_out();
965 Node* addp = shift->unique_out();
966 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
967 Node* mem = addp->last_out(j);
968 if (UseCondCardMark && mem->is_Load()) {
969 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
970 // The load is checking if the card has been written so
971 // replace it with zero to fold the test.
972 macro->replace_node(mem, macro->intcon(0));
973 continue;
974 }
975 assert(mem->is_Store(), "store required");
976 macro->replace_node(mem, mem->in(MemNode::Memory));
977 }
978 }
979 }
980
981 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
982 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
983 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
984 c = c->unique_ctrl_out();
985 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
986 c = c->unique_ctrl_out();
987 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
988 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
989 assert(iff->is_If(), "expect test");
990 if (!is_shenandoah_marking_if(igvn, iff)) {
991 c = c->unique_ctrl_out();
992 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
993 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
994 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
995 }
996 Node* cmpx = iff->in(1)->in(1);
997 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
998 igvn->rehash_node_delayed(call);
999 call->del_req(call->req()-1);
1000 }
1001
1002 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
1003 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
1004 igvn->add_users_to_worklist(node);
1005 }
1006 }
1007
1008 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
1009 for (uint i = 0; i < useful.size(); i++) {
1010 Node* n = useful.at(i);
1011 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
1012 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1013 C->record_for_igvn(n->fast_out(i));
1014 }
1015 }
1016 }
1017
1018 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
1019 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
1020 if (!useful.member(n)) {
1021 state()->remove_load_reference_barrier(n);
1022 }
1023 }
1024 }
1025
1026 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
1027 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
1028 }
1029
1030 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
1031 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
1032 }
1033
1034 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
1035 // expanded later, then now is the time to do so.
1036 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
1037
1038 #ifdef ASSERT
1039 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
1040 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) {
1041 ShenandoahBarrierC2Support::verify(Compile::current()->root());
1042 } else if (phase == BarrierSetC2::BeforeCodeGen) {
1043 // Verify Shenandoah pre-barriers
1044 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset());
1045
1046 Unique_Node_List visited;
1047 Node_List worklist;
1048 // We're going to walk control flow backwards starting from the Root
1049 worklist.push(compile->root());
1050 while (worklist.size() > 0) {
1051 Node *x = worklist.pop();
1052 if (x == nullptr || x == compile->top()) continue;
1053 if (visited.member(x)) {
1054 continue;
1055 } else {
1056 visited.push(x);
1057 }
1058
1059 if (x->is_Region()) {
1060 for (uint i = 1; i < x->req(); i++) {
1061 worklist.push(x->in(i));
1062 }
1063 } else {
1064 worklist.push(x->in(0));
1065 // We are looking for the pattern:
1066 // /->ThreadLocal
1067 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
1068 // \->ConI(0)
1069 // We want to verify that the If and the LoadB have the same control
1070 // See GraphKit::g1_write_barrier_pre()
1071 if (x->is_If()) {
1072 IfNode *iff = x->as_If();
1073 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
1074 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
1075 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
1076 && cmp->in(1)->is_Load()) {
1077 LoadNode *load = cmp->in(1)->as_Load();
1078 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
1079 && load->in(2)->in(3)->is_Con()
1080 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
1081
1082 Node *if_ctrl = iff->in(0);
1083 Node *load_ctrl = load->in(0);
1084
1085 if (if_ctrl != load_ctrl) {
1086 // Skip possible CProj->NeverBranch in infinite loops
1087 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
1088 && if_ctrl->in(0)->is_NeverBranch()) {
1089 if_ctrl = if_ctrl->in(0)->in(0);
1090 }
1091 }
1092 assert(load_ctrl != nullptr && if_ctrl == load_ctrl, "controls must match");
1093 }
1094 }
1095 }
1096 }
1097 }
1098 }
1099 }
1100 }
1101 #endif
1102
1103 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
1104 if (is_shenandoah_wb_pre_call(n)) {
1105 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1106 if (n->req() > cnt) {
1107 Node* addp = n->in(cnt);
1108 if (has_only_shenandoah_wb_pre_uses(addp)) {
1109 n->del_req(cnt);
1110 if (can_reshape) {
1111 phase->is_IterGVN()->_worklist.push(addp);
1112 }
1113 return n;
1114 }
1115 }
1116 }
1117 if (n->Opcode() == Op_CmpP) {
1118 Node* in1 = n->in(1);
1119 Node* in2 = n->in(2);
1120
1121 // If one input is null, then step over the strong LRB barriers on the other input
1122 if (in1->bottom_type() == TypePtr::NULL_PTR &&
1123 !((in2->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1124 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in2)->decorators()))) {
1125 in2 = step_over_gc_barrier(in2);
1126 }
1127 if (in2->bottom_type() == TypePtr::NULL_PTR &&
1128 !((in1->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1129 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in1)->decorators()))) {
1130 in1 = step_over_gc_barrier(in1);
1131 }
1132
1133 if (in1 != n->in(1)) {
1134 n->set_req_X(1, in1, phase);
1135 assert(in2 == n->in(2), "only one change");
1136 return n;
1137 }
1138 if (in2 != n->in(2)) {
1139 n->set_req_X(2, in2, phase);
1140 return n;
1141 }
1142 } else if (can_reshape &&
1143 n->Opcode() == Op_If &&
1144 ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1145 n->in(0) != nullptr &&
1146 n->outcnt() == 2) {
1147 Node* dom = n->in(0);
1148 Node* prev_dom = n;
1149 int op = n->Opcode();
1150 int dist = 16;
1151 // Search up the dominator tree for another heap stable test
1152 while (dom->Opcode() != op || // Not same opcode?
1153 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1?
1154 prev_dom->in(0) != dom) { // One path of test does not dominate?
1155 if (dist < 0) return nullptr;
1156
1157 dist--;
1158 prev_dom = dom;
1159 dom = IfNode::up_one_dom(dom);
1160 if (!dom) return nullptr;
1161 }
1162
1163 // Check that we did not follow a loop back to ourselves
1164 if (n == dom) {
1165 return nullptr;
1166 }
1167
1168 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN(), false);
1169 }
1170
1171 return nullptr;
1172 }
1173
1174 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1175 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1176 Node* u = n->fast_out(i);
1177 if (!is_shenandoah_wb_pre_call(u)) {
1178 return false;
1179 }
1180 }
1181 return n->outcnt() > 0;
1182 }
1183
1184 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode, Unique_Node_List& dead_nodes) const {
1185 switch (opcode) {
1186 case Op_CallLeaf:
1187 case Op_CallLeafNoFP: {
1188 assert (n->is_Call(), "");
1189 CallNode *call = n->as_Call();
1190 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1191 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1192 if (call->req() > cnt) {
1193 assert(call->req() == cnt + 1, "only one extra input");
1194 Node *addp = call->in(cnt);
1195 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1196 call->del_req(cnt);
1197 }
1198 }
1199 return false;
1200 }
1201 case Op_ShenandoahCompareAndSwapP:
1202 case Op_ShenandoahCompareAndSwapN:
1203 case Op_ShenandoahWeakCompareAndSwapN:
1204 case Op_ShenandoahWeakCompareAndSwapP:
1205 case Op_ShenandoahCompareAndExchangeP:
1206 case Op_ShenandoahCompareAndExchangeN:
1207 return true;
1208 case Op_ShenandoahLoadReferenceBarrier:
1209 assert(false, "should have been expanded already");
1210 return true;
1211 default:
1212 return false;
1213 }
1214 }
1215
1216 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1217 switch (opcode) {
1218 case Op_ShenandoahCompareAndExchangeP:
1219 case Op_ShenandoahCompareAndExchangeN:
1220 conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1221 // fallthrough
1222 case Op_ShenandoahWeakCompareAndSwapP:
1223 case Op_ShenandoahWeakCompareAndSwapN:
1224 case Op_ShenandoahCompareAndSwapP:
1225 case Op_ShenandoahCompareAndSwapN:
1226 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1227 return true;
1228 case Op_StoreP: {
1229 Node* adr = n->in(MemNode::Address);
1230 const Type* adr_type = gvn->type(adr);
1231 // Pointer stores in Shenandoah barriers looks like unsafe access.
1232 // Ignore such stores to be able scalar replace non-escaping
1233 // allocations.
1234 if (adr_type->isa_rawptr() && adr->is_AddP()) {
1235 Node* base = conn_graph->get_addp_base(adr);
1236 if (base->Opcode() == Op_LoadP &&
1237 base->in(MemNode::Address)->is_AddP()) {
1238 adr = base->in(MemNode::Address);
1239 Node* tls = conn_graph->get_addp_base(adr);
1240 if (tls->Opcode() == Op_ThreadLocal) {
1241 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1242 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1243 if (offs == buf_offset) {
1244 return true; // Pre barrier previous oop value store.
1245 }
1246 }
1247 }
1248 }
1249 return false;
1250 }
1251 case Op_ShenandoahLoadReferenceBarrier:
1252 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist);
1253 return true;
1254 default:
1255 // Nothing
1256 break;
1257 }
1258 return false;
1259 }
1260
1261 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1262 switch (opcode) {
1263 case Op_ShenandoahCompareAndExchangeP:
1264 case Op_ShenandoahCompareAndExchangeN: {
1265 Node *adr = n->in(MemNode::Address);
1266 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, nullptr);
1267 // fallthrough
1268 }
1269 case Op_ShenandoahCompareAndSwapP:
1270 case Op_ShenandoahCompareAndSwapN:
1271 case Op_ShenandoahWeakCompareAndSwapP:
1272 case Op_ShenandoahWeakCompareAndSwapN:
1273 return conn_graph->add_final_edges_unsafe_access(n, opcode);
1274 case Op_ShenandoahLoadReferenceBarrier:
1275 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), nullptr);
1276 return true;
1277 default:
1278 // Nothing
1279 break;
1280 }
1281 return false;
1282 }
1283
1284 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1285 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1286 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1287
1288 }
1289
1290 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1291 switch (opcode) {
1292 case Op_ShenandoahCompareAndExchangeP:
1293 case Op_ShenandoahCompareAndExchangeN:
1294 case Op_ShenandoahWeakCompareAndSwapP:
1295 case Op_ShenandoahWeakCompareAndSwapN:
1296 case Op_ShenandoahCompareAndSwapP:
1297 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree
1298 Node* newval = n->in(MemNode::ValueIn);
1299 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1300 Node* pair = new BinaryNode(oldval, newval);
1301 n->set_req(MemNode::ValueIn,pair);
1302 n->del_req(LoadStoreConditionalNode::ExpectedIn);
1303 return true;
1304 }
1305 default:
1306 break;
1307 }
1308 return false;
1309 }
1310
1311 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1312 return xop == Op_ShenandoahCompareAndExchangeP ||
1313 xop == Op_ShenandoahCompareAndExchangeN ||
1314 xop == Op_ShenandoahWeakCompareAndSwapP ||
1315 xop == Op_ShenandoahWeakCompareAndSwapN ||
1316 xop == Op_ShenandoahCompareAndSwapN ||
1317 xop == Op_ShenandoahCompareAndSwapP;
1318 }