1 /*
2 * Copyright (c) 2018, 2026, 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 "classfile/javaClasses.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
29 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
30 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
31 #include "gc/shenandoah/shenandoahBarrierSet.hpp"
32 #include "gc/shenandoah/shenandoahCardTable.hpp"
33 #include "gc/shenandoah/shenandoahForwarding.hpp"
34 #include "gc/shenandoah/shenandoahHeap.hpp"
35 #include "gc/shenandoah/shenandoahRuntime.hpp"
36 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
37 #include "opto/arraycopynode.hpp"
38 #include "opto/escape.hpp"
39 #include "opto/graphKit.hpp"
40 #include "opto/idealKit.hpp"
41 #include "opto/macro.hpp"
42 #include "opto/movenode.hpp"
43 #include "opto/narrowptrnode.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46
47 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() {
48 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2());
49 }
50
51 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena)
52 : _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8, 0, nullptr)) {
53 }
54
55 int ShenandoahBarrierSetC2State::load_reference_barriers_count() const {
56 return _load_reference_barriers->length();
57 }
58
59 ShenandoahLoadReferenceBarrierNode* ShenandoahBarrierSetC2State::load_reference_barrier(int idx) const {
60 return _load_reference_barriers->at(idx);
61 }
62
63 void ShenandoahBarrierSetC2State::add_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
64 assert(!_load_reference_barriers->contains(n), "duplicate entry in barrier list");
65 _load_reference_barriers->append(n);
66 }
67
68 void ShenandoahBarrierSetC2State::remove_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
69 if (_load_reference_barriers->contains(n)) {
70 _load_reference_barriers->remove(n);
71 }
72 }
73
74 #define __ kit->
75
76 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseValues* phase, Node* adr,
77 BasicType bt, uint adr_idx) const {
78 intptr_t offset = 0;
79 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
80 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
81
82 if (offset == Type::OffsetBot) {
83 return false; // cannot unalias unless there are precise offsets
84 }
85
86 if (alloc == nullptr) {
87 return false; // No allocation found
88 }
89
90 intptr_t size_in_bytes = type2aelembytes(bt);
91
92 Node* mem = __ memory(adr_idx); // start searching here...
93
94 for (int cnt = 0; cnt < 50; cnt++) {
95
96 if (mem->is_Store()) {
97
98 Node* st_adr = mem->in(MemNode::Address);
99 intptr_t st_offset = 0;
100 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
101
102 if (st_base == nullptr) {
103 break; // inscrutable pointer
104 }
105
106 // Break we have found a store with same base and offset as ours so break
107 if (st_base == base && st_offset == offset) {
108 break;
109 }
110
111 if (st_offset != offset && st_offset != Type::OffsetBot) {
112 const int MAX_STORE = BytesPerLong;
113 if (st_offset >= offset + size_in_bytes ||
114 st_offset <= offset - MAX_STORE ||
115 st_offset <= offset - mem->as_Store()->memory_size()) {
116 // Success: The offsets are provably independent.
117 // (You may ask, why not just test st_offset != offset and be done?
118 // The answer is that stores of different sizes can co-exist
119 // in the same sequence of RawMem effects. We sometimes initialize
120 // a whole 'tile' of array elements with a single jint or jlong.)
121 mem = mem->in(MemNode::Memory);
122 continue; // advance through independent store memory
123 }
124 }
125
126 if (st_base != base
127 && MemNode::detect_ptr_independence(base, alloc, st_base,
128 AllocateNode::Ideal_allocation(st_base),
129 phase)) {
130 // Success: The bases are provably independent.
131 mem = mem->in(MemNode::Memory);
132 continue; // advance through independent store memory
133 }
134 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
135
136 InitializeNode* st_init = mem->in(0)->as_Initialize();
137 AllocateNode* st_alloc = st_init->allocation();
138
139 // Make sure that we are looking at the same allocation site.
140 // The alloc variable is guaranteed to not be null here from earlier check.
141 if (alloc == st_alloc) {
142 // Check that the initialization is storing null so that no previous store
143 // has been moved up and directly write a reference
144 Node* captured_store = st_init->find_captured_store(offset,
145 type2aelembytes(T_OBJECT),
146 phase);
147 if (captured_store == nullptr || captured_store == st_init->zero_memory()) {
148 return true;
149 }
150 }
151 }
152
153 // Unless there is an explicit 'continue', we must bail out here,
154 // because 'mem' is an inscrutable memory state (e.g., a call).
155 break;
156 }
157
158 return false;
159 }
160
161 #undef __
162 #define __ ideal.
163
164 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit,
165 bool do_load,
166 Node* obj,
167 Node* adr,
168 uint alias_idx,
169 Node* val,
170 const TypeOopPtr* val_type,
171 Node* pre_val,
172 BasicType bt) const {
173 // Some sanity checks
174 // Note: val is unused in this routine.
175
176 if (do_load) {
177 // We need to generate the load of the previous value
178 assert(adr != nullptr, "where are loading from?");
179 assert(pre_val == nullptr, "loaded already?");
180 assert(val_type != nullptr, "need a type");
181
182 if (ReduceInitialCardMarks
183 && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
184 return;
185 }
186
187 } else {
188 // In this case both val_type and alias_idx are unused.
189 assert(pre_val != nullptr, "must be loaded already");
190 // Nothing to be done if pre_val is null.
191 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
192 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
193 }
194 assert(bt == T_OBJECT, "or we shouldn't be here");
195
196 IdealKit ideal(kit, true);
197
198 Node* tls = __ thread(); // ThreadLocalStorage
199
200 Node* no_base = __ top();
201 Node* zero = __ ConI(0);
202 Node* zeroX = __ ConX(0);
203
204 float likely = PROB_LIKELY(0.999);
205 float unlikely = PROB_UNLIKELY(0.999);
206
207 // Offsets into the thread
208 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
209 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
210
211 // Now the actual pointers into the thread
212 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
213 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
214
215 // Now some of the values
216 Node* marking;
217 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())));
218 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
219 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
220 assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape");
221
222 // if (!marking)
223 __ if_then(marking, BoolTest::ne, zero, unlikely); {
224 BasicType index_bt = TypeX_X->basic_type();
225 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading Shenandoah SATBMarkQueue::_index with wrong size.");
226 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
227
228 if (do_load) {
229 // load original value
230 // alias_idx correct??
231 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
232 }
233
234 // if (pre_val != nullptr)
235 __ if_then(pre_val, BoolTest::ne, kit->null()); {
236 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
237
238 // is the queue for this thread full?
239 __ if_then(index, BoolTest::ne, zeroX, likely); {
240
241 // decrement the index
242 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
243
244 // Now get the buffer location we will log the previous value into and store it
245 Node *log_addr = __ AddP(no_base, buffer, next_index);
246 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
247 // update the index
248 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
249
250 } __ else_(); {
251
252 // logging buffer is full, call the runtime
253 const TypeFunc *tf = ShenandoahBarrierSetC2::write_barrier_pre_Type();
254 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre), "shenandoah_wb_pre", pre_val);
255 } __ end_if(); // (!index)
256 } __ end_if(); // (pre_val != nullptr)
257 } __ end_if(); // (!marking)
258
259 // Final sync IdealKit and GraphKit.
260 kit->final_sync(ideal);
261
262 if (ShenandoahSATBBarrier && adr != nullptr) {
263 Node* c = kit->control();
264 Node* call = c->in(1)->in(1)->in(1)->in(0);
265 assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected");
266 call->add_req(adr);
267 }
268 }
269
270 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) {
271 return call->is_CallLeaf() &&
272 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre);
273 }
274
275 bool ShenandoahBarrierSetC2::is_shenandoah_clone_call(Node* call) {
276 return call->is_CallLeaf() &&
277 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::clone_barrier);
278 }
279
280 bool ShenandoahBarrierSetC2::is_shenandoah_lrb_call(Node* call) {
281 if (!call->is_CallLeaf()) {
282 return false;
283 }
284
285 address entry_point = call->as_CallLeaf()->entry_point();
286 return (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong)) ||
287 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow)) ||
288 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak)) ||
289 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow)) ||
290 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom)) ||
291 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow));
292 }
293
294 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseValues* phase, Node* n) {
295 if (n->Opcode() != Op_If) {
296 return false;
297 }
298
299 Node* bol = n->in(1);
300 assert(bol->is_Bool(), "");
301 Node* cmpx = bol->in(1);
302 if (bol->as_Bool()->_test._test == BoolTest::ne &&
303 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
304 is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
305 cmpx->in(1)->in(2)->is_Con() &&
306 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
307 return true;
308 }
309
310 return false;
311 }
312
313 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
314 if (!n->is_Load()) return false;
315 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
316 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
317 && n->in(2)->in(3)->is_Con()
318 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
319 }
320
321 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
322 bool do_load,
323 Node* obj,
324 Node* adr,
325 uint alias_idx,
326 Node* val,
327 const TypeOopPtr* val_type,
328 Node* pre_val,
329 BasicType bt) const {
330 if (ShenandoahSATBBarrier) {
331 IdealKit ideal(kit);
332 kit->sync_kit(ideal);
333
334 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
335
336 ideal.sync_kit(kit);
337 kit->final_sync(ideal);
338 }
339 }
340
341 // Helper that guards and inserts a pre-barrier.
342 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
343 Node* pre_val, bool need_mem_bar) const {
344 // We could be accessing the referent field of a reference object. If so, when Shenandoah
345 // is enabled, we need to log the value in the referent field in an SATB buffer.
346 // This routine performs some compile time filters and generates suitable
347 // runtime filters that guard the pre-barrier code.
348 // Also add memory barrier for non volatile load from the referent field
349 // to prevent commoning of loads across safepoint.
350
351 // Some compile time checks.
352
353 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
354 const TypeX* otype = offset->find_intptr_t_type();
355 if (otype != nullptr && otype->is_con() &&
356 otype->get_con() != java_lang_ref_Reference::referent_offset()) {
357 // Constant offset but not the reference_offset so just return
358 return;
359 }
360
361 // We only need to generate the runtime guards for instances.
362 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
363 if (btype != nullptr) {
364 if (btype->isa_aryptr()) {
365 // Array type so nothing to do
366 return;
367 }
368
369 const TypeInstPtr* itype = btype->isa_instptr();
370 if (itype != nullptr) {
371 // Can the klass of base_oop be statically determined to be
372 // _not_ a sub-class of Reference and _not_ Object?
373 ciKlass* klass = itype->instance_klass();
374 if (klass->is_loaded() &&
375 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
376 !kit->env()->Object_klass()->is_subtype_of(klass)) {
377 return;
378 }
379 }
380 }
381
382 // The compile time filters did not reject base_oop/offset so
383 // we need to generate the following runtime filters
384 //
385 // if (offset == java_lang_ref_Reference::_reference_offset) {
386 // if (instance_of(base, java.lang.ref.Reference)) {
387 // pre_barrier(_, pre_val, ...);
388 // }
389 // }
390
391 float likely = PROB_LIKELY( 0.999);
392 float unlikely = PROB_UNLIKELY(0.999);
393
394 IdealKit ideal(kit);
395
396 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset());
397
398 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
399 // Update graphKit memory and control from IdealKit.
400 kit->sync_kit(ideal);
401
402 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
403 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
404
405 // Update IdealKit memory and control from graphKit.
406 __ sync_kit(kit);
407
408 Node* one = __ ConI(1);
409 // is_instof == 0 if base_oop == nullptr
410 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
411
412 // Update graphKit from IdeakKit.
413 kit->sync_kit(ideal);
414
415 // Use the pre-barrier to record the value in the referent field
416 satb_write_barrier_pre(kit, false /* do_load */,
417 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */,
418 pre_val /* pre_val */,
419 T_OBJECT);
420 if (need_mem_bar) {
421 // Add memory barrier to prevent commoning reads from this field
422 // across safepoint since GC can change its value.
423 kit->insert_mem_bar(Op_MemBarCPUOrder);
424 }
425 // Update IdealKit from graphKit.
426 __ sync_kit(kit);
427
428 } __ end_if(); // _ref_type != ref_none
429 } __ end_if(); // offset == referent_offset
430
431 // Final sync IdealKit and GraphKit.
432 kit->final_sync(ideal);
433 }
434
435 void ShenandoahBarrierSetC2::post_barrier(GraphKit* kit,
436 Node* ctl,
437 Node* oop_store,
438 Node* obj,
439 Node* adr,
440 uint adr_idx,
441 Node* val,
442 BasicType bt,
443 bool use_precise) const {
444 assert(ShenandoahCardBarrier, "Should have been checked by caller");
445
446 // No store check needed if we're storing a null.
447 if (val != nullptr && val->is_Con()) {
448 // must be either an oop or null
449 const Type* t = val->bottom_type();
450 if (t == TypePtr::NULL_PTR || t == Type::TOP)
451 return;
452 }
453
454 if (ReduceInitialCardMarks && obj == kit->just_allocated_object(kit->control())) {
455 // We use card marks to track old to young references in Generational Shenandoah;
456 // see flag ShenandoahCardBarrier above.
457 // Objects are always allocated in the young generation and initialized
458 // before they are promoted. There's always a safepoint (e.g. at final mark)
459 // before an object is promoted from young to old. Promotion entails dirtying of
460 // the cards backing promoted objects, so they will be guaranteed to be scanned
461 // at the next remembered set scan of the old generation.
462 // Thus, we can safely skip card-marking of initializing stores on a
463 // freshly-allocated object. If any of the assumptions above change in
464 // the future, this code will need to be re-examined; see check in
465 // ShenandoahCardBarrier::on_slowpath_allocation_exit().
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 Node* tls = __ thread(); // ThreadLocalStorage
479
480 // Convert the pointer to an int prior to doing math on it
481 Node* cast = __ CastPX(__ ctrl(), adr);
482
483 Node* curr_ct_holder_offset = __ ConX(in_bytes(ShenandoahThreadLocalData::card_table_offset()));
484 Node* curr_ct_holder_addr = __ AddP(__ top(), tls, curr_ct_holder_offset);
485 Node* curr_ct_base_addr = __ load( __ ctrl(), curr_ct_holder_addr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
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(), curr_ct_base_addr, 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_barrier_pre_Type = nullptr;
523 const TypeFunc* ShenandoahBarrierSetC2::_clone_barrier_Type = nullptr;
524 const TypeFunc* ShenandoahBarrierSetC2::_load_reference_barrier_Type = nullptr;
525
526 inline const TypeFunc* ShenandoahBarrierSetC2::write_barrier_pre_Type() {
527 assert(ShenandoahBarrierSetC2::_write_barrier_pre_Type != nullptr, "should be initialized");
528 return ShenandoahBarrierSetC2::_write_barrier_pre_Type;
529 }
530
531 inline const TypeFunc* ShenandoahBarrierSetC2::clone_barrier_Type() {
532 assert(ShenandoahBarrierSetC2::_clone_barrier_Type != nullptr, "should be initialized");
533 return ShenandoahBarrierSetC2::_clone_barrier_Type;
534 }
535
536 const TypeFunc* ShenandoahBarrierSetC2::load_reference_barrier_Type() {
537 assert(ShenandoahBarrierSetC2::_load_reference_barrier_Type != nullptr, "should be initialized");
538 return ShenandoahBarrierSetC2::_load_reference_barrier_Type;
539 }
540
541 void ShenandoahBarrierSetC2::init() {
542 ShenandoahBarrierSetC2::make_write_barrier_pre_Type();
543 ShenandoahBarrierSetC2::make_clone_barrier_Type();
544 ShenandoahBarrierSetC2::make_load_reference_barrier_Type();
545 }
546
547 void ShenandoahBarrierSetC2::make_write_barrier_pre_Type() {
548 assert(ShenandoahBarrierSetC2::_write_barrier_pre_Type == nullptr, "should be");
549 const Type **fields = TypeTuple::fields(1);
550 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
551 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
552
553 // create result type (range)
554 fields = TypeTuple::fields(0);
555 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
556
557 ShenandoahBarrierSetC2::_write_barrier_pre_Type = TypeFunc::make(domain, range);
558 }
559
560 void ShenandoahBarrierSetC2::make_clone_barrier_Type() {
561 assert(ShenandoahBarrierSetC2::_clone_barrier_Type == nullptr, "should be");
562 const Type **fields = TypeTuple::fields(1);
563 fields[TypeFunc::Parms+0] = TypeOopPtr::NOTNULL; // src oop
564 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
565
566 // create result type (range)
567 fields = TypeTuple::fields(0);
568 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
569
570 ShenandoahBarrierSetC2::_clone_barrier_Type = TypeFunc::make(domain, range);
571 }
572
573 void ShenandoahBarrierSetC2::make_load_reference_barrier_Type() {
574 assert(ShenandoahBarrierSetC2::_load_reference_barrier_Type == nullptr, "should be");
575 const Type **fields = TypeTuple::fields(2);
576 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM; // original field value
577 fields[TypeFunc::Parms+1] = TypeRawPtr::BOTTOM; // original load address
578
579 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
580
581 // create result type (range)
582 fields = TypeTuple::fields(1);
583 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM;
584 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
585
586 ShenandoahBarrierSetC2::_load_reference_barrier_Type = TypeFunc::make(domain, range);
587 }
588
589 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
590 DecoratorSet decorators = access.decorators();
591
592 const TypePtr* adr_type = access.addr().type();
593 Node* adr = access.addr().node();
594
595 bool no_keepalive = (decorators & AS_NO_KEEPALIVE) != 0;
596
597 if (!access.is_oop()) {
598 return BarrierSetC2::store_at_resolved(access, val);
599 }
600
601 if (no_keepalive) {
602 // No keep-alive means no need for the pre-barrier.
603 return BarrierSetC2::store_at_resolved(access, val);
604 }
605
606 if (access.is_parse_access()) {
607 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
608 GraphKit* kit = parse_access.kit();
609
610 uint adr_idx = kit->C->get_alias_index(adr_type);
611 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
612 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
613 static_cast<const TypeOopPtr*>(val.type()), nullptr /* pre_val */, access.type());
614
615 Node* result = BarrierSetC2::store_at_resolved(access, val);
616
617 if (ShenandoahCardBarrier) {
618 const bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
619 const bool is_array = (decorators & IS_ARRAY) != 0;
620 const bool use_precise = is_array || anonymous;
621 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
622 adr, adr_idx, val.node(), access.type(), use_precise);
623 }
624 return result;
625 } else {
626 assert(access.is_opt_access(), "only for optimization passes");
627 assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code");
628 return BarrierSetC2::store_at_resolved(access, val);
629 }
630 }
631
632 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
633 // 1: non-reference load, no additional barrier is needed
634 if (!access.is_oop()) {
635 return BarrierSetC2::load_at_resolved(access, val_type);
636 }
637
638 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
639 DecoratorSet decorators = access.decorators();
640 BasicType type = access.type();
641
642 // 2: apply LRB if needed
643 if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
644 load = new ShenandoahLoadReferenceBarrierNode(nullptr, load, decorators);
645 if (access.is_parse_access()) {
646 load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load);
647 } else {
648 load = static_cast<C2OptAccess &>(access).gvn().transform(load);
649 }
650 }
651
652 // 3: apply keep-alive barrier for java.lang.ref.Reference if needed
653 if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
654 Node* top = Compile::current()->top();
655 Node* adr = access.addr().node();
656 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
657 Node* obj = access.base();
658
659 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
660 bool on_weak_ref = (decorators & (ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF)) != 0;
661 bool keep_alive = (decorators & AS_NO_KEEPALIVE) == 0;
662
663 // If we are reading the value of the referent field of a Reference
664 // object (either by using Unsafe directly or through reflection)
665 // then, if SATB is enabled, we need to record the referent in an
666 // SATB log buffer using the pre-barrier mechanism.
667 // Also we need to add memory barrier to prevent commoning reads
668 // from this field across safepoint since GC can change its value.
669 if (!on_weak_ref || (unknown && (offset == top || obj == top)) || !keep_alive) {
670 return load;
671 }
672
673 assert(access.is_parse_access(), "entry not supported at optimization time");
674 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
675 GraphKit* kit = parse_access.kit();
676 bool mismatched = (decorators & C2_MISMATCHED) != 0;
677 bool is_unordered = (decorators & MO_UNORDERED) != 0;
678 bool in_native = (decorators & IN_NATIVE) != 0;
679 bool need_cpu_mem_bar = !is_unordered || mismatched || in_native;
680
681 if (on_weak_ref) {
682 // Use the pre-barrier to record the value in the referent field
683 satb_write_barrier_pre(kit, false /* do_load */,
684 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */,
685 load /* pre_val */, T_OBJECT);
686 // Add memory barrier to prevent commoning reads from this field
687 // across safepoint since GC can change its value.
688 kit->insert_mem_bar(Op_MemBarCPUOrder);
689 } else if (unknown) {
690 // We do not require a mem bar inside pre_barrier if need_mem_bar
691 // is set: the barriers would be emitted by us.
692 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
693 }
694 }
695
696 return load;
697 }
698
699 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
700 Node* new_val, const Type* value_type) const {
701 GraphKit* kit = access.kit();
702 if (access.is_oop()) {
703 shenandoah_write_barrier_pre(kit, false /* do_load */,
704 nullptr, nullptr, max_juint, nullptr, nullptr,
705 expected_val /* pre_val */, T_OBJECT);
706
707 MemNode::MemOrd mo = access.mem_node_mo();
708 Node* mem = access.memory();
709 Node* adr = access.addr().node();
710 const TypePtr* adr_type = access.addr().type();
711 Node* load_store = nullptr;
712
713 #ifdef _LP64
714 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
715 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
716 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
717 if (ShenandoahCASBarrier) {
718 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
719 } else {
720 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
721 }
722 } else
723 #endif
724 {
725 if (ShenandoahCASBarrier) {
726 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
727 } else {
728 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
729 }
730 }
731
732 access.set_raw_access(load_store);
733 pin_atomic_op(access);
734
735 #ifdef _LP64
736 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
737 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
738 }
739 #endif
740 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, load_store, access.decorators()));
741 if (ShenandoahCardBarrier) {
742 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
743 access.addr().node(), access.alias_idx(), new_val, T_OBJECT, true);
744 }
745 return load_store;
746 }
747 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
748 }
749
750 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
751 Node* new_val, const Type* value_type) const {
752 GraphKit* kit = access.kit();
753 if (access.is_oop()) {
754 shenandoah_write_barrier_pre(kit, false /* do_load */,
755 nullptr, nullptr, max_juint, nullptr, nullptr,
756 expected_val /* pre_val */, T_OBJECT);
757 DecoratorSet decorators = access.decorators();
758 MemNode::MemOrd mo = access.mem_node_mo();
759 Node* mem = access.memory();
760 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
761 Node* load_store = nullptr;
762 Node* adr = access.addr().node();
763 #ifdef _LP64
764 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
765 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
766 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
767 if (ShenandoahCASBarrier) {
768 if (is_weak_cas) {
769 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
770 } else {
771 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
772 }
773 } else {
774 if (is_weak_cas) {
775 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
776 } else {
777 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
778 }
779 }
780 } else
781 #endif
782 {
783 if (ShenandoahCASBarrier) {
784 if (is_weak_cas) {
785 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
786 } else {
787 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
788 }
789 } else {
790 if (is_weak_cas) {
791 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
792 } else {
793 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
794 }
795 }
796 }
797 access.set_raw_access(load_store);
798 pin_atomic_op(access);
799 if (ShenandoahCardBarrier) {
800 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
801 access.addr().node(), access.alias_idx(), new_val, T_OBJECT, true);
802 }
803 return load_store;
804 }
805 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
806 }
807
808 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
809 GraphKit* kit = access.kit();
810 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
811 if (access.is_oop()) {
812 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, result, access.decorators()));
813 shenandoah_write_barrier_pre(kit, false /* do_load */,
814 nullptr, nullptr, max_juint, nullptr, nullptr,
815 result /* pre_val */, T_OBJECT);
816 if (ShenandoahCardBarrier) {
817 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
818 access.addr().node(), access.alias_idx(), val, T_OBJECT, true);
819 }
820 }
821 return result;
822 }
823
824
825 bool ShenandoahBarrierSetC2::is_gc_pre_barrier_node(Node* node) const {
826 return is_shenandoah_wb_pre_call(node);
827 }
828
829 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
830 return (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) ||
831 is_shenandoah_lrb_call(node) ||
832 is_shenandoah_wb_pre_call(node) ||
833 is_shenandoah_clone_call(node);
834 }
835
836 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
837 if (c == nullptr) {
838 return c;
839 }
840 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
841 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
842 }
843 return c;
844 }
845
846 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
847 return !ShenandoahBarrierC2Support::expand(C, igvn);
848 }
849
850 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
851 if (mode == LoopOptsShenandoahExpand) {
852 assert(UseShenandoahGC, "only for shenandoah");
853 ShenandoahBarrierC2Support::pin_and_expand(phase);
854 return true;
855 }
856 return false;
857 }
858
859 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, bool is_clone_instance, ArrayCopyPhase phase) const {
860 bool is_oop = is_reference_type(type);
861 if (!is_oop) {
862 return false;
863 }
864 if (ShenandoahSATBBarrier && tightly_coupled_alloc) {
865 if (phase == Optimization) {
866 return false;
867 }
868 return !is_clone;
869 }
870 return true;
871 }
872
873 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
874 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
875 if (src_type->isa_instptr() != nullptr) {
876 ciInstanceKlass* ik = src_type->is_instptr()->instance_klass();
877 if ((src_type->klass_is_exact() || !ik->has_subklass()) && !ik->has_injected_fields()) {
878 if (ik->has_object_fields()) {
879 return true;
880 } else {
881 if (!src_type->klass_is_exact()) {
882 Compile::current()->dependencies()->assert_leaf_type(ik);
883 }
884 }
885 } else {
886 return true;
887 }
888 } else if (src_type->isa_aryptr()) {
889 BasicType src_elem = src_type->isa_aryptr()->elem()->array_element_basic_type();
890 if (is_reference_type(src_elem, true)) {
891 return true;
892 }
893 } else {
894 return true;
895 }
896 return false;
897 }
898
899 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
900 Node* ctrl = ac->in(TypeFunc::Control);
901 Node* mem = ac->in(TypeFunc::Memory);
902 Node* src_base = ac->in(ArrayCopyNode::Src);
903 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
904 Node* dest_base = ac->in(ArrayCopyNode::Dest);
905 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
906 Node* length = ac->in(ArrayCopyNode::Length);
907
908 Node* src = phase->basic_plus_adr(src_base, src_offset);
909 Node* dest = phase->basic_plus_adr(dest_base, dest_offset);
910
911 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
912 // Check if heap is has forwarded objects. If it does, we need to call into the special
913 // routine that would fix up source references before we can continue.
914
915 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
916 Node* region = new RegionNode(PATH_LIMIT);
917 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM);
918
919 Node* thread = phase->transform_later(new ThreadLocalNode());
920 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
921 Node* gc_state_addr = phase->transform_later(AddPNode::make_off_heap(thread, offset));
922
923 uint gc_state_idx = Compile::AliasIdxRaw;
924 const TypePtr* gc_state_adr_type = nullptr; // debug-mode-only argument
925 DEBUG_ONLY(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
926
927 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered));
928 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED)));
929 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT)));
930 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne));
931
932 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If();
933 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff));
934 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff));
935
936 // Heap is stable, no need to do anything additional
937 region->init_req(_heap_stable, stable_ctrl);
938 mem_phi->init_req(_heap_stable, mem);
939
940 // Heap is unstable, call into clone barrier stub
941 Node* call = phase->make_leaf_call(unstable_ctrl, mem,
942 ShenandoahBarrierSetC2::clone_barrier_Type(),
943 CAST_FROM_FN_PTR(address, ShenandoahRuntime::clone_barrier),
944 "shenandoah_clone",
945 TypeRawPtr::BOTTOM,
946 src_base);
947 call = phase->transform_later(call);
948
949 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control));
950 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory));
951 region->init_req(_heap_unstable, ctrl);
952 mem_phi->init_req(_heap_unstable, mem);
953
954 // Wire up the actual arraycopy stub now
955 ctrl = phase->transform_later(region);
956 mem = phase->transform_later(mem_phi);
957
958 const char* name = "arraycopy";
959 call = phase->make_leaf_call(ctrl, mem,
960 OptoRuntime::fast_arraycopy_Type(),
961 phase->basictype2arraycopy(T_LONG, nullptr, nullptr, true, name, true),
962 name, TypeRawPtr::BOTTOM,
963 src, dest, length
964 LP64_ONLY(COMMA phase->top()));
965 call = phase->transform_later(call);
966
967 // Hook up the whole thing into the graph
968 phase->replace_node(ac, call);
969 } else {
970 BarrierSetC2::clone_at_expansion(phase, ac);
971 }
972 }
973
974
975 // Support for macro expanded GC barriers
976 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
977 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
978 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
979 }
980 }
981
982 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
983 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
984 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
985 }
986 }
987
988 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
989 if (is_shenandoah_wb_pre_call(node)) {
990 shenandoah_eliminate_wb_pre(node, igvn);
991 }
992 if (ShenandoahCardBarrier && node->Opcode() == Op_CastP2X) {
993 Node* shift = node->unique_out();
994 Node* addp = shift->unique_out();
995 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
996 Node* mem = addp->last_out(j);
997 if (UseCondCardMark && mem->is_Load()) {
998 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
999 // The load is checking if the card has been written so
1000 // replace it with zero to fold the test.
1001 igvn->replace_node(mem, igvn->intcon(0));
1002 continue;
1003 }
1004 assert(mem->is_Store(), "store required");
1005 igvn->replace_node(mem, mem->in(MemNode::Memory));
1006 }
1007 }
1008 }
1009
1010 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
1011 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
1012 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
1013 c = c->unique_ctrl_out();
1014 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
1015 c = c->unique_ctrl_out();
1016 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
1017 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
1018 assert(iff->is_If(), "expect test");
1019 if (!is_shenandoah_marking_if(igvn, iff)) {
1020 c = c->unique_ctrl_out();
1021 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
1022 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
1023 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
1024 }
1025 Node* cmpx = iff->in(1)->in(1);
1026 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
1027 igvn->rehash_node_delayed(call);
1028 call->del_req(call->req()-1);
1029 }
1030
1031 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
1032 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
1033 igvn->add_users_to_worklist(node);
1034 }
1035 }
1036
1037 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
1038 for (uint i = 0; i < useful.size(); i++) {
1039 Node* n = useful.at(i);
1040 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
1041 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1042 C->record_for_igvn(n->fast_out(i));
1043 }
1044 }
1045 }
1046
1047 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
1048 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
1049 if (!useful.member(n)) {
1050 state()->remove_load_reference_barrier(n);
1051 }
1052 }
1053 }
1054
1055 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
1056 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
1057 }
1058
1059 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
1060 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
1061 }
1062
1063 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
1064 // expanded later, then now is the time to do so.
1065 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
1066
1067 #ifdef ASSERT
1068 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
1069 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) {
1070 ShenandoahBarrierC2Support::verify(Compile::current()->root());
1071 } else if (phase == BarrierSetC2::BeforeCodeGen) {
1072 // Verify Shenandoah pre-barriers
1073 const int gc_state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
1074
1075 Unique_Node_List visited;
1076 Node_List worklist;
1077 // We're going to walk control flow backwards starting from the Root
1078 worklist.push(compile->root());
1079 while (worklist.size() > 0) {
1080 Node *x = worklist.pop();
1081 if (x == nullptr || x == compile->top()) {
1082 continue;
1083 }
1084
1085 if (visited.member(x)) {
1086 continue;
1087 } else {
1088 visited.push(x);
1089 }
1090
1091 if (x->is_Region()) {
1092 for (uint i = 1; i < x->req(); i++) {
1093 worklist.push(x->in(i));
1094 }
1095 } else {
1096 worklist.push(x->in(0));
1097 // We are looking for the pattern:
1098 // /->ThreadLocal
1099 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
1100 // \->ConI(0)
1101 // We want to verify that the If and the LoadB have the same control
1102 // See GraphKit::g1_write_barrier_pre()
1103 if (x->is_If()) {
1104 IfNode *iff = x->as_If();
1105 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
1106 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
1107 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
1108 && cmp->in(1)->is_Load()) {
1109 LoadNode *load = cmp->in(1)->as_Load();
1110 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
1111 && load->in(2)->in(3)->is_Con()
1112 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == gc_state_offset) {
1113
1114 Node *if_ctrl = iff->in(0);
1115 Node *load_ctrl = load->in(0);
1116
1117 if (if_ctrl != load_ctrl) {
1118 // Skip possible CProj->NeverBranch in infinite loops
1119 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
1120 && if_ctrl->in(0)->is_NeverBranch()) {
1121 if_ctrl = if_ctrl->in(0)->in(0);
1122 }
1123 }
1124 assert(load_ctrl != nullptr && if_ctrl == load_ctrl, "controls must match");
1125 }
1126 }
1127 }
1128 }
1129 }
1130 }
1131 }
1132 }
1133 #endif
1134
1135 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
1136 if (is_shenandoah_wb_pre_call(n)) {
1137 uint cnt = ShenandoahBarrierSetC2::write_barrier_pre_Type()->domain_sig()->cnt();
1138 if (n->req() > cnt) {
1139 Node* addp = n->in(cnt);
1140 if (has_only_shenandoah_wb_pre_uses(addp)) {
1141 n->del_req(cnt);
1142 if (can_reshape) {
1143 phase->is_IterGVN()->_worklist.push(addp);
1144 }
1145 return n;
1146 }
1147 }
1148 }
1149 if (n->Opcode() == Op_CmpP) {
1150 Node* in1 = n->in(1);
1151 Node* in2 = n->in(2);
1152
1153 // If one input is null, then step over the strong LRB barriers on the other input
1154 if (in1->bottom_type() == TypePtr::NULL_PTR &&
1155 !((in2->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1156 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in2)->decorators()))) {
1157 in2 = step_over_gc_barrier(in2);
1158 }
1159 if (in2->bottom_type() == TypePtr::NULL_PTR &&
1160 !((in1->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1161 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in1)->decorators()))) {
1162 in1 = step_over_gc_barrier(in1);
1163 }
1164
1165 if (in1 != n->in(1)) {
1166 n->set_req_X(1, in1, phase);
1167 assert(in2 == n->in(2), "only one change");
1168 return n;
1169 }
1170 if (in2 != n->in(2)) {
1171 n->set_req_X(2, in2, phase);
1172 return n;
1173 }
1174 } else if (can_reshape &&
1175 n->Opcode() == Op_If &&
1176 ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1177 n->in(0) != nullptr &&
1178 n->outcnt() == 2) {
1179 Node* dom = n->in(0);
1180 Node* prev_dom = n;
1181 int op = n->Opcode();
1182 int dist = 16;
1183 // Search up the dominator tree for another heap stable test
1184 while (dom->Opcode() != op || // Not same opcode?
1185 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1?
1186 prev_dom->in(0) != dom) { // One path of test does not dominate?
1187 if (dist < 0) return nullptr;
1188
1189 dist--;
1190 prev_dom = dom;
1191 dom = IfNode::up_one_dom(dom);
1192 if (!dom) return nullptr;
1193 }
1194
1195 // Check that we did not follow a loop back to ourselves
1196 if (n == dom) {
1197 return nullptr;
1198 }
1199
1200 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN(), false);
1201 }
1202
1203 return nullptr;
1204 }
1205
1206 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1207 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1208 Node* u = n->fast_out(i);
1209 if (!is_shenandoah_wb_pre_call(u)) {
1210 return false;
1211 }
1212 }
1213 return n->outcnt() > 0;
1214 }
1215
1216 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode, Unique_Node_List& dead_nodes) const {
1217 switch (opcode) {
1218 case Op_CallLeaf:
1219 case Op_CallLeafNoFP: {
1220 assert (n->is_Call(), "");
1221 CallNode *call = n->as_Call();
1222 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1223 uint cnt = ShenandoahBarrierSetC2::write_barrier_pre_Type()->domain_sig()->cnt();
1224 if (call->req() > cnt) {
1225 assert(call->req() == cnt + 1, "only one extra input");
1226 Node *addp = call->in(cnt);
1227 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1228 call->del_req(cnt);
1229 }
1230 }
1231 return false;
1232 }
1233 case Op_ShenandoahCompareAndSwapP:
1234 case Op_ShenandoahCompareAndSwapN:
1235 case Op_ShenandoahWeakCompareAndSwapN:
1236 case Op_ShenandoahWeakCompareAndSwapP:
1237 case Op_ShenandoahCompareAndExchangeP:
1238 case Op_ShenandoahCompareAndExchangeN:
1239 return true;
1240 case Op_ShenandoahLoadReferenceBarrier:
1241 assert(false, "should have been expanded already");
1242 return true;
1243 default:
1244 return false;
1245 }
1246 }
1247
1248 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1249 switch (opcode) {
1250 case Op_ShenandoahCompareAndExchangeP:
1251 case Op_ShenandoahCompareAndExchangeN:
1252 conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1253 // fallthrough
1254 case Op_ShenandoahWeakCompareAndSwapP:
1255 case Op_ShenandoahWeakCompareAndSwapN:
1256 case Op_ShenandoahCompareAndSwapP:
1257 case Op_ShenandoahCompareAndSwapN:
1258 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1259 return true;
1260 case Op_StoreP: {
1261 Node* adr = n->in(MemNode::Address);
1262 const Type* adr_type = gvn->type(adr);
1263 // Pointer stores in Shenandoah barriers looks like unsafe access.
1264 // Ignore such stores to be able scalar replace non-escaping
1265 // allocations.
1266 if (adr_type->isa_rawptr() && adr->is_AddP()) {
1267 Node* base = conn_graph->get_addp_base(adr);
1268 if (base->Opcode() == Op_LoadP &&
1269 base->in(MemNode::Address)->is_AddP()) {
1270 adr = base->in(MemNode::Address);
1271 Node* tls = conn_graph->get_addp_base(adr);
1272 if (tls->Opcode() == Op_ThreadLocal) {
1273 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1274 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1275 if (offs == buf_offset) {
1276 return true; // Pre barrier previous oop value store.
1277 }
1278 }
1279 }
1280 }
1281 return false;
1282 }
1283 case Op_ShenandoahLoadReferenceBarrier:
1284 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist);
1285 return true;
1286 default:
1287 // Nothing
1288 break;
1289 }
1290 return false;
1291 }
1292
1293 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1294 switch (opcode) {
1295 case Op_ShenandoahCompareAndExchangeP:
1296 case Op_ShenandoahCompareAndExchangeN: {
1297 Node *adr = n->in(MemNode::Address);
1298 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, nullptr);
1299 // fallthrough
1300 }
1301 case Op_ShenandoahCompareAndSwapP:
1302 case Op_ShenandoahCompareAndSwapN:
1303 case Op_ShenandoahWeakCompareAndSwapP:
1304 case Op_ShenandoahWeakCompareAndSwapN:
1305 return conn_graph->add_final_edges_unsafe_access(n, opcode);
1306 case Op_ShenandoahLoadReferenceBarrier:
1307 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), nullptr);
1308 return true;
1309 default:
1310 // Nothing
1311 break;
1312 }
1313 return false;
1314 }
1315
1316 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1317 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1318 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1319
1320 }
1321
1322 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1323 switch (opcode) {
1324 case Op_ShenandoahCompareAndExchangeP:
1325 case Op_ShenandoahCompareAndExchangeN:
1326 case Op_ShenandoahWeakCompareAndSwapP:
1327 case Op_ShenandoahWeakCompareAndSwapN:
1328 case Op_ShenandoahCompareAndSwapP:
1329 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree
1330 Node* newval = n->in(MemNode::ValueIn);
1331 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1332 Node* pair = new BinaryNode(oldval, newval);
1333 n->set_req(MemNode::ValueIn,pair);
1334 n->del_req(LoadStoreConditionalNode::ExpectedIn);
1335 return true;
1336 }
1337 default:
1338 break;
1339 }
1340 return false;
1341 }
1342
1343 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1344 return xop == Op_ShenandoahCompareAndExchangeP ||
1345 xop == Op_ShenandoahCompareAndExchangeN ||
1346 xop == Op_ShenandoahWeakCompareAndSwapP ||
1347 xop == Op_ShenandoahWeakCompareAndSwapN ||
1348 xop == Op_ShenandoahCompareAndSwapN ||
1349 xop == Op_ShenandoahCompareAndSwapP;
1350 }