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 "classfile/javaClasses.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shenandoah/shenandoahBarrierSet.hpp"
29 #include "gc/shenandoah/shenandoahCardTable.hpp"
30 #include "gc/shenandoah/shenandoahForwarding.hpp"
31 #include "gc/shenandoah/shenandoahHeap.hpp"
32 #include "gc/shenandoah/shenandoahRuntime.hpp"
33 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
34 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
35 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
36 #include "gc/shenandoah/heuristics/shenandoahHeuristics.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_ref_field_pre_Type();
254 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre), "shenandoah_wb_pre",
255 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);
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::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 use card marks to track old to young references in Generational Shenandoah;
457 // see flag ShenandoahCardBarrier above.
458 // Objects are always allocated in the young generation and initialized
459 // before they are promoted. There's always a safepoint (e.g. at final mark)
460 // before an object is promoted from young to old. Promotion entails dirtying of
461 // the cards backing promoted objects, so they will be guaranteed to be scanned
462 // at the next remembered set scan of the old generation.
463 // Thus, we can safely skip card-marking of initializing stores on a
464 // freshly-allocated object. If any of the assumptions above change in
465 // the future, this code will need to be re-examined; see check in
466 // ShenandoahCardBarrier::on_slowpath_allocation_exit().
467 return;
468 }
469
470 if (!use_precise) {
471 // All card marks for a (non-array) instance are in one place:
472 adr = obj;
473 }
474 // (Else it's an array (or unknown), and we want more precise card marks.)
475 assert(adr != nullptr, "");
476
477 IdealKit ideal(kit, true);
478
479 Node* tls = __ thread(); // ThreadLocalStorage
480
481 // Convert the pointer to an int prior to doing math on it
482 Node* cast = __ CastPX(__ ctrl(), adr);
483
484 Node* curr_ct_holder_offset = __ ConX(in_bytes(ShenandoahThreadLocalData::card_table_offset()));
485 Node* curr_ct_holder_addr = __ AddP(__ top(), tls, curr_ct_holder_offset);
486 Node* curr_ct_base_addr = __ load( __ ctrl(), curr_ct_holder_addr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
487
488 // Divide by card size
489 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift()) );
490
491 // Combine card table base and card offset
492 Node* card_adr = __ AddP(__ top(), curr_ct_base_addr, card_offset);
493
494 // Get the alias_index for raw card-mark memory
495 int adr_type = Compile::AliasIdxRaw;
496 Node* zero = __ ConI(0); // Dirty card value
497
498 if (UseCondCardMark) {
499 // The classic GC reference write barrier is typically implemented
500 // as a store into the global card mark table. Unfortunately
501 // unconditional stores can result in false sharing and excessive
502 // coherence traffic as well as false transactional aborts.
503 // UseCondCardMark enables MP "polite" conditional card mark
504 // stores. In theory we could relax the load from ctrl() to
505 // no_ctrl, but that doesn't buy much latitude.
506 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, T_BYTE, adr_type);
507 __ if_then(card_val, BoolTest::ne, zero);
508 }
509
510 // Smash zero into card
511 __ store(__ ctrl(), card_adr, zero, T_BYTE, adr_type, MemNode::unordered);
512
513 if (UseCondCardMark) {
514 __ end_if();
515 }
516
517 // Final sync IdealKit and GraphKit.
518 kit->final_sync(ideal);
519 }
520
521 #undef __
522
523 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_Type() {
524 const Type **fields = TypeTuple::fields(2);
525 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
526 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
527 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
528
529 // create result type (range)
530 fields = TypeTuple::fields(0);
531 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
532
533 return TypeFunc::make(domain, range);
534 }
535
536 const TypeFunc* ShenandoahBarrierSetC2::clone_barrier_Type() {
537 const Type **fields = TypeTuple::fields(1);
538 fields[TypeFunc::Parms+0] = TypeOopPtr::NOTNULL; // src oop
539 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
540
541 // create result type (range)
542 fields = TypeTuple::fields(0);
543 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
544
545 return TypeFunc::make(domain, range);
546 }
547
548 const TypeFunc* ShenandoahBarrierSetC2::load_reference_barrier_Type() {
549 const Type **fields = TypeTuple::fields(2);
550 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM; // original field value
551 fields[TypeFunc::Parms+1] = TypeRawPtr::BOTTOM; // original load address
552
553 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
554
555 // create result type (range)
556 fields = TypeTuple::fields(1);
557 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM;
558 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
559
560 return TypeFunc::make(domain, range);
561 }
562
563 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
564 DecoratorSet decorators = access.decorators();
565
566 const TypePtr* adr_type = access.addr().type();
567 Node* adr = access.addr().node();
568
569 bool no_keepalive = (decorators & AS_NO_KEEPALIVE) != 0;
570
571 if (!access.is_oop()) {
572 return BarrierSetC2::store_at_resolved(access, val);
573 }
574
575 if (no_keepalive) {
576 // No keep-alive means no need for the pre-barrier.
577 return BarrierSetC2::store_at_resolved(access, val);
578 }
579
580 if (access.is_parse_access()) {
581 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
582 GraphKit* kit = parse_access.kit();
583
584 uint adr_idx = kit->C->get_alias_index(adr_type);
585 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
586 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
587 static_cast<const TypeOopPtr*>(val.type()), nullptr /* pre_val */, access.type());
588
589 Node* result = BarrierSetC2::store_at_resolved(access, val);
590
591 if (ShenandoahCardBarrier) {
592 const bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
593 const bool is_array = (decorators & IS_ARRAY) != 0;
594 const bool use_precise = is_array || anonymous;
595 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
596 adr, adr_idx, val.node(), access.type(), use_precise);
597 }
598 return result;
599 } else {
600 assert(access.is_opt_access(), "only for optimization passes");
601 assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code");
602 return BarrierSetC2::store_at_resolved(access, val);
603 }
604 }
605
606 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
607 // 1: non-reference load, no additional barrier is needed
608 if (!access.is_oop()) {
609 return BarrierSetC2::load_at_resolved(access, val_type);
610 }
611
612 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
613 DecoratorSet decorators = access.decorators();
614 BasicType type = access.type();
615
616 // 2: apply LRB if needed
617 if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
618 load = new ShenandoahLoadReferenceBarrierNode(nullptr, load, decorators);
619 if (access.is_parse_access()) {
620 load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load);
621 } else {
622 load = static_cast<C2OptAccess &>(access).gvn().transform(load);
623 }
624 }
625
626 // 3: apply keep-alive barrier for java.lang.ref.Reference if needed
627 if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
628 Node* top = Compile::current()->top();
629 Node* adr = access.addr().node();
630 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
631 Node* obj = access.base();
632
633 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
634 bool on_weak_ref = (decorators & (ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF)) != 0;
635 bool keep_alive = (decorators & AS_NO_KEEPALIVE) == 0;
636
637 // If we are reading the value of the referent field of a Reference
638 // object (either by using Unsafe directly or through reflection)
639 // then, if SATB is enabled, we need to record the referent in an
640 // SATB log buffer using the pre-barrier mechanism.
641 // Also we need to add memory barrier to prevent commoning reads
642 // from this field across safepoint since GC can change its value.
643 if (!on_weak_ref || (unknown && (offset == top || obj == top)) || !keep_alive) {
644 return load;
645 }
646
647 assert(access.is_parse_access(), "entry not supported at optimization time");
648 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
649 GraphKit* kit = parse_access.kit();
650 bool mismatched = (decorators & C2_MISMATCHED) != 0;
651 bool is_unordered = (decorators & MO_UNORDERED) != 0;
652 bool in_native = (decorators & IN_NATIVE) != 0;
653 bool need_cpu_mem_bar = !is_unordered || mismatched || in_native;
654
655 if (on_weak_ref) {
656 // Use the pre-barrier to record the value in the referent field
657 satb_write_barrier_pre(kit, false /* do_load */,
658 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */,
659 load /* pre_val */, T_OBJECT);
660 // Add memory barrier to prevent commoning reads from this field
661 // across safepoint since GC can change its value.
662 kit->insert_mem_bar(Op_MemBarCPUOrder);
663 } else if (unknown) {
664 // We do not require a mem bar inside pre_barrier if need_mem_bar
665 // is set: the barriers would be emitted by us.
666 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
667 }
668 }
669
670 return load;
671 }
672
673 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
674 Node* new_val, const Type* value_type) const {
675 GraphKit* kit = access.kit();
676 if (access.is_oop()) {
677 shenandoah_write_barrier_pre(kit, false /* do_load */,
678 nullptr, nullptr, max_juint, nullptr, nullptr,
679 expected_val /* pre_val */, T_OBJECT);
680
681 MemNode::MemOrd mo = access.mem_node_mo();
682 Node* mem = access.memory();
683 Node* adr = access.addr().node();
684 const TypePtr* adr_type = access.addr().type();
685 Node* load_store = nullptr;
686
687 #ifdef _LP64
688 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
689 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
690 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
691 if (ShenandoahCASBarrier) {
692 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
693 } else {
694 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
695 }
696 } else
697 #endif
698 {
699 if (ShenandoahCASBarrier) {
700 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
701 } else {
702 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
703 }
704 }
705
706 access.set_raw_access(load_store);
707 pin_atomic_op(access);
708
709 #ifdef _LP64
710 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
711 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
712 }
713 #endif
714 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, load_store, access.decorators()));
715 if (ShenandoahCardBarrier) {
716 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
717 access.addr().node(), access.alias_idx(), new_val, T_OBJECT, true);
718 }
719 return load_store;
720 }
721 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
722 }
723
724 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
725 Node* new_val, const Type* value_type) const {
726 GraphKit* kit = access.kit();
727 if (access.is_oop()) {
728 shenandoah_write_barrier_pre(kit, false /* do_load */,
729 nullptr, nullptr, max_juint, nullptr, nullptr,
730 expected_val /* pre_val */, T_OBJECT);
731 DecoratorSet decorators = access.decorators();
732 MemNode::MemOrd mo = access.mem_node_mo();
733 Node* mem = access.memory();
734 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
735 Node* load_store = nullptr;
736 Node* adr = access.addr().node();
737 #ifdef _LP64
738 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
739 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
740 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
741 if (ShenandoahCASBarrier) {
742 if (is_weak_cas) {
743 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
744 } else {
745 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
746 }
747 } else {
748 if (is_weak_cas) {
749 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
750 } else {
751 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
752 }
753 }
754 } else
755 #endif
756 {
757 if (ShenandoahCASBarrier) {
758 if (is_weak_cas) {
759 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
760 } else {
761 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
762 }
763 } else {
764 if (is_weak_cas) {
765 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
766 } else {
767 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
768 }
769 }
770 }
771 access.set_raw_access(load_store);
772 pin_atomic_op(access);
773 if (ShenandoahCardBarrier) {
774 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
775 access.addr().node(), access.alias_idx(), new_val, T_OBJECT, true);
776 }
777 return load_store;
778 }
779 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
780 }
781
782 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
783 GraphKit* kit = access.kit();
784 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
785 if (access.is_oop()) {
786 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, result, access.decorators()));
787 shenandoah_write_barrier_pre(kit, false /* do_load */,
788 nullptr, nullptr, max_juint, nullptr, nullptr,
789 result /* pre_val */, T_OBJECT);
790 if (ShenandoahCardBarrier) {
791 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
792 access.addr().node(), access.alias_idx(), val, T_OBJECT, true);
793 }
794 }
795 return result;
796 }
797
798
799 bool ShenandoahBarrierSetC2::is_gc_pre_barrier_node(Node* node) const {
800 return is_shenandoah_wb_pre_call(node);
801 }
802
803 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
804 return (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) ||
805 is_shenandoah_lrb_call(node) ||
806 is_shenandoah_wb_pre_call(node) ||
807 is_shenandoah_clone_call(node);
808 }
809
810 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
811 if (c == nullptr) {
812 return c;
813 }
814 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
815 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
816 }
817 return c;
818 }
819
820 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
821 return !ShenandoahBarrierC2Support::expand(C, igvn);
822 }
823
824 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
825 if (mode == LoopOptsShenandoahExpand) {
826 assert(UseShenandoahGC, "only for shenandoah");
827 ShenandoahBarrierC2Support::pin_and_expand(phase);
828 return true;
829 } else if (mode == LoopOptsShenandoahPostExpand) {
830 assert(UseShenandoahGC, "only for shenandoah");
831 visited.clear();
832 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
833 return true;
834 }
835 return false;
836 }
837
838 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, bool is_clone_instance, ArrayCopyPhase phase) const {
839 bool is_oop = is_reference_type(type);
840 if (!is_oop) {
841 return false;
842 }
843 if (ShenandoahSATBBarrier && tightly_coupled_alloc) {
844 if (phase == Optimization) {
845 return false;
846 }
847 return !is_clone;
848 }
849 return true;
850 }
851
852 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
853 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
854 if (src_type->isa_instptr() != nullptr) {
855 ciInstanceKlass* ik = src_type->is_instptr()->instance_klass();
856 if ((src_type->klass_is_exact() || !ik->has_subklass()) && !ik->has_injected_fields()) {
857 if (ik->has_object_fields()) {
858 return true;
859 } else {
860 if (!src_type->klass_is_exact()) {
861 Compile::current()->dependencies()->assert_leaf_type(ik);
862 }
863 }
864 } else {
865 return true;
866 }
867 } else if (src_type->isa_aryptr()) {
868 BasicType src_elem = src_type->isa_aryptr()->elem()->array_element_basic_type();
869 if (is_reference_type(src_elem, true)) {
870 return true;
871 }
872 } else {
873 return true;
874 }
875 return false;
876 }
877
878 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
879 Node* ctrl = ac->in(TypeFunc::Control);
880 Node* mem = ac->in(TypeFunc::Memory);
881 Node* src_base = ac->in(ArrayCopyNode::Src);
882 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
883 Node* dest_base = ac->in(ArrayCopyNode::Dest);
884 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
885 Node* length = ac->in(ArrayCopyNode::Length);
886
887 Node* src = phase->basic_plus_adr(src_base, src_offset);
888 Node* dest = phase->basic_plus_adr(dest_base, dest_offset);
889
890 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
891 // Check if heap is has forwarded objects. If it does, we need to call into the special
892 // routine that would fix up source references before we can continue.
893
894 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
895 Node* region = new RegionNode(PATH_LIMIT);
896 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM);
897
898 Node* thread = phase->transform_later(new ThreadLocalNode());
899 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
900 Node* gc_state_addr = phase->transform_later(new AddPNode(phase->C->top(), thread, offset));
901
902 uint gc_state_idx = Compile::AliasIdxRaw;
903 const TypePtr* gc_state_adr_type = nullptr; // debug-mode-only argument
904 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
905
906 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered));
907 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED)));
908 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT)));
909 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne));
910
911 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If();
912 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff));
913 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff));
914
915 // Heap is stable, no need to do anything additional
916 region->init_req(_heap_stable, stable_ctrl);
917 mem_phi->init_req(_heap_stable, mem);
918
919 // Heap is unstable, call into clone barrier stub
920 Node* call = phase->make_leaf_call(unstable_ctrl, mem,
921 ShenandoahBarrierSetC2::clone_barrier_Type(),
922 CAST_FROM_FN_PTR(address, ShenandoahRuntime::clone_barrier),
923 "shenandoah_clone",
924 TypeRawPtr::BOTTOM,
925 src_base);
926 call = phase->transform_later(call);
927
928 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control));
929 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory));
930 region->init_req(_heap_unstable, ctrl);
931 mem_phi->init_req(_heap_unstable, mem);
932
933 // Wire up the actual arraycopy stub now
934 ctrl = phase->transform_later(region);
935 mem = phase->transform_later(mem_phi);
936
937 const char* name = "arraycopy";
938 call = phase->make_leaf_call(ctrl, mem,
939 OptoRuntime::fast_arraycopy_Type(),
940 phase->basictype2arraycopy(T_LONG, nullptr, nullptr, true, name, true),
941 name, TypeRawPtr::BOTTOM,
942 src, dest, length
943 LP64_ONLY(COMMA phase->top()));
944 call = phase->transform_later(call);
945
946 // Hook up the whole thing into the graph
947 phase->igvn().replace_node(ac, call);
948 } else {
949 BarrierSetC2::clone_at_expansion(phase, ac);
950 }
951 }
952
953
954 // Support for macro expanded GC barriers
955 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
956 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
957 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
958 }
959 }
960
961 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
962 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
963 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
964 }
965 }
966
967 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const {
968 if (is_shenandoah_wb_pre_call(node)) {
969 shenandoah_eliminate_wb_pre(node, ¯o->igvn());
970 }
971 if (ShenandoahCardBarrier && node->Opcode() == Op_CastP2X) {
972 Node* shift = node->unique_out();
973 Node* addp = shift->unique_out();
974 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
975 Node* mem = addp->last_out(j);
976 if (UseCondCardMark && mem->is_Load()) {
977 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
978 // The load is checking if the card has been written so
979 // replace it with zero to fold the test.
980 macro->replace_node(mem, macro->intcon(0));
981 continue;
982 }
983 assert(mem->is_Store(), "store required");
984 macro->replace_node(mem, mem->in(MemNode::Memory));
985 }
986 }
987 }
988
989 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
990 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
991 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
992 c = c->unique_ctrl_out();
993 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
994 c = c->unique_ctrl_out();
995 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
996 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
997 assert(iff->is_If(), "expect test");
998 if (!is_shenandoah_marking_if(igvn, iff)) {
999 c = c->unique_ctrl_out();
1000 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
1001 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
1002 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
1003 }
1004 Node* cmpx = iff->in(1)->in(1);
1005 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
1006 igvn->rehash_node_delayed(call);
1007 call->del_req(call->req()-1);
1008 }
1009
1010 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
1011 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
1012 igvn->add_users_to_worklist(node);
1013 }
1014 }
1015
1016 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
1017 for (uint i = 0; i < useful.size(); i++) {
1018 Node* n = useful.at(i);
1019 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
1020 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1021 C->record_for_igvn(n->fast_out(i));
1022 }
1023 }
1024 }
1025
1026 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
1027 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
1028 if (!useful.member(n)) {
1029 state()->remove_load_reference_barrier(n);
1030 }
1031 }
1032 }
1033
1034 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
1035 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
1036 }
1037
1038 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
1039 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
1040 }
1041
1042 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
1043 // expanded later, then now is the time to do so.
1044 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
1045
1046 #ifdef ASSERT
1047 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
1048 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) {
1049 ShenandoahBarrierC2Support::verify(Compile::current()->root());
1050 } else if (phase == BarrierSetC2::BeforeCodeGen) {
1051 // Verify Shenandoah pre-barriers
1052 const int gc_state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
1053
1054 Unique_Node_List visited;
1055 Node_List worklist;
1056 // We're going to walk control flow backwards starting from the Root
1057 worklist.push(compile->root());
1058 while (worklist.size() > 0) {
1059 Node *x = worklist.pop();
1060 if (x == nullptr || x == compile->top()) {
1061 continue;
1062 }
1063
1064 if (visited.member(x)) {
1065 continue;
1066 } else {
1067 visited.push(x);
1068 }
1069
1070 if (x->is_Region()) {
1071 for (uint i = 1; i < x->req(); i++) {
1072 worklist.push(x->in(i));
1073 }
1074 } else {
1075 worklist.push(x->in(0));
1076 // We are looking for the pattern:
1077 // /->ThreadLocal
1078 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
1079 // \->ConI(0)
1080 // We want to verify that the If and the LoadB have the same control
1081 // See GraphKit::g1_write_barrier_pre()
1082 if (x->is_If()) {
1083 IfNode *iff = x->as_If();
1084 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
1085 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
1086 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
1087 && cmp->in(1)->is_Load()) {
1088 LoadNode *load = cmp->in(1)->as_Load();
1089 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
1090 && load->in(2)->in(3)->is_Con()
1091 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == gc_state_offset) {
1092
1093 Node *if_ctrl = iff->in(0);
1094 Node *load_ctrl = load->in(0);
1095
1096 if (if_ctrl != load_ctrl) {
1097 // Skip possible CProj->NeverBranch in infinite loops
1098 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
1099 && if_ctrl->in(0)->is_NeverBranch()) {
1100 if_ctrl = if_ctrl->in(0)->in(0);
1101 }
1102 }
1103 assert(load_ctrl != nullptr && if_ctrl == load_ctrl, "controls must match");
1104 }
1105 }
1106 }
1107 }
1108 }
1109 }
1110 }
1111 }
1112 #endif
1113
1114 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
1115 if (is_shenandoah_wb_pre_call(n)) {
1116 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_Type()->domain()->cnt();
1117 if (n->req() > cnt) {
1118 Node* addp = n->in(cnt);
1119 if (has_only_shenandoah_wb_pre_uses(addp)) {
1120 n->del_req(cnt);
1121 if (can_reshape) {
1122 phase->is_IterGVN()->_worklist.push(addp);
1123 }
1124 return n;
1125 }
1126 }
1127 }
1128 if (n->Opcode() == Op_CmpP) {
1129 Node* in1 = n->in(1);
1130 Node* in2 = n->in(2);
1131
1132 // If one input is null, then step over the strong LRB barriers on the other input
1133 if (in1->bottom_type() == TypePtr::NULL_PTR &&
1134 !((in2->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1135 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in2)->decorators()))) {
1136 in2 = step_over_gc_barrier(in2);
1137 }
1138 if (in2->bottom_type() == TypePtr::NULL_PTR &&
1139 !((in1->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1140 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in1)->decorators()))) {
1141 in1 = step_over_gc_barrier(in1);
1142 }
1143
1144 if (in1 != n->in(1)) {
1145 n->set_req_X(1, in1, phase);
1146 assert(in2 == n->in(2), "only one change");
1147 return n;
1148 }
1149 if (in2 != n->in(2)) {
1150 n->set_req_X(2, in2, phase);
1151 return n;
1152 }
1153 } else if (can_reshape &&
1154 n->Opcode() == Op_If &&
1155 ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1156 n->in(0) != nullptr &&
1157 n->outcnt() == 2) {
1158 Node* dom = n->in(0);
1159 Node* prev_dom = n;
1160 int op = n->Opcode();
1161 int dist = 16;
1162 // Search up the dominator tree for another heap stable test
1163 while (dom->Opcode() != op || // Not same opcode?
1164 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1?
1165 prev_dom->in(0) != dom) { // One path of test does not dominate?
1166 if (dist < 0) return nullptr;
1167
1168 dist--;
1169 prev_dom = dom;
1170 dom = IfNode::up_one_dom(dom);
1171 if (!dom) return nullptr;
1172 }
1173
1174 // Check that we did not follow a loop back to ourselves
1175 if (n == dom) {
1176 return nullptr;
1177 }
1178
1179 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN(), false);
1180 }
1181
1182 return nullptr;
1183 }
1184
1185 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1186 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1187 Node* u = n->fast_out(i);
1188 if (!is_shenandoah_wb_pre_call(u)) {
1189 return false;
1190 }
1191 }
1192 return n->outcnt() > 0;
1193 }
1194
1195 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode, Unique_Node_List& dead_nodes) const {
1196 switch (opcode) {
1197 case Op_CallLeaf:
1198 case Op_CallLeafNoFP: {
1199 assert (n->is_Call(), "");
1200 CallNode *call = n->as_Call();
1201 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1202 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_Type()->domain()->cnt();
1203 if (call->req() > cnt) {
1204 assert(call->req() == cnt + 1, "only one extra input");
1205 Node *addp = call->in(cnt);
1206 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1207 call->del_req(cnt);
1208 }
1209 }
1210 return false;
1211 }
1212 case Op_ShenandoahCompareAndSwapP:
1213 case Op_ShenandoahCompareAndSwapN:
1214 case Op_ShenandoahWeakCompareAndSwapN:
1215 case Op_ShenandoahWeakCompareAndSwapP:
1216 case Op_ShenandoahCompareAndExchangeP:
1217 case Op_ShenandoahCompareAndExchangeN:
1218 return true;
1219 case Op_ShenandoahLoadReferenceBarrier:
1220 assert(false, "should have been expanded already");
1221 return true;
1222 default:
1223 return false;
1224 }
1225 }
1226
1227 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1228 switch (opcode) {
1229 case Op_ShenandoahCompareAndExchangeP:
1230 case Op_ShenandoahCompareAndExchangeN:
1231 conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1232 // fallthrough
1233 case Op_ShenandoahWeakCompareAndSwapP:
1234 case Op_ShenandoahWeakCompareAndSwapN:
1235 case Op_ShenandoahCompareAndSwapP:
1236 case Op_ShenandoahCompareAndSwapN:
1237 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1238 return true;
1239 case Op_StoreP: {
1240 Node* adr = n->in(MemNode::Address);
1241 const Type* adr_type = gvn->type(adr);
1242 // Pointer stores in Shenandoah barriers looks like unsafe access.
1243 // Ignore such stores to be able scalar replace non-escaping
1244 // allocations.
1245 if (adr_type->isa_rawptr() && adr->is_AddP()) {
1246 Node* base = conn_graph->get_addp_base(adr);
1247 if (base->Opcode() == Op_LoadP &&
1248 base->in(MemNode::Address)->is_AddP()) {
1249 adr = base->in(MemNode::Address);
1250 Node* tls = conn_graph->get_addp_base(adr);
1251 if (tls->Opcode() == Op_ThreadLocal) {
1252 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1253 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1254 if (offs == buf_offset) {
1255 return true; // Pre barrier previous oop value store.
1256 }
1257 }
1258 }
1259 }
1260 return false;
1261 }
1262 case Op_ShenandoahLoadReferenceBarrier:
1263 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist);
1264 return true;
1265 default:
1266 // Nothing
1267 break;
1268 }
1269 return false;
1270 }
1271
1272 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1273 switch (opcode) {
1274 case Op_ShenandoahCompareAndExchangeP:
1275 case Op_ShenandoahCompareAndExchangeN: {
1276 Node *adr = n->in(MemNode::Address);
1277 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, nullptr);
1278 // fallthrough
1279 }
1280 case Op_ShenandoahCompareAndSwapP:
1281 case Op_ShenandoahCompareAndSwapN:
1282 case Op_ShenandoahWeakCompareAndSwapP:
1283 case Op_ShenandoahWeakCompareAndSwapN:
1284 return conn_graph->add_final_edges_unsafe_access(n, opcode);
1285 case Op_ShenandoahLoadReferenceBarrier:
1286 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), nullptr);
1287 return true;
1288 default:
1289 // Nothing
1290 break;
1291 }
1292 return false;
1293 }
1294
1295 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1296 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1297 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1298
1299 }
1300
1301 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1302 switch (opcode) {
1303 case Op_ShenandoahCompareAndExchangeP:
1304 case Op_ShenandoahCompareAndExchangeN:
1305 case Op_ShenandoahWeakCompareAndSwapP:
1306 case Op_ShenandoahWeakCompareAndSwapN:
1307 case Op_ShenandoahCompareAndSwapP:
1308 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree
1309 Node* newval = n->in(MemNode::ValueIn);
1310 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1311 Node* pair = new BinaryNode(oldval, newval);
1312 n->set_req(MemNode::ValueIn,pair);
1313 n->del_req(LoadStoreConditionalNode::ExpectedIn);
1314 return true;
1315 }
1316 default:
1317 break;
1318 }
1319 return false;
1320 }
1321
1322 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1323 return xop == Op_ShenandoahCompareAndExchangeP ||
1324 xop == Op_ShenandoahCompareAndExchangeN ||
1325 xop == Op_ShenandoahWeakCompareAndSwapP ||
1326 xop == Op_ShenandoahWeakCompareAndSwapN ||
1327 xop == Op_ShenandoahCompareAndSwapN ||
1328 xop == Op_ShenandoahCompareAndSwapP;
1329 }