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