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 : _iu_barriers(new (comp_arena) GrowableArray<ShenandoahIUBarrierNode*>(comp_arena, 8, 0, nullptr)),
54 _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8, 0, nullptr)) {
55 }
56
57 int ShenandoahBarrierSetC2State::iu_barriers_count() const {
58 return _iu_barriers->length();
59 }
60
61 ShenandoahIUBarrierNode* ShenandoahBarrierSetC2State::iu_barrier(int idx) const {
62 return _iu_barriers->at(idx);
63 }
64
65 void ShenandoahBarrierSetC2State::add_iu_barrier(ShenandoahIUBarrierNode* n) {
66 assert(!_iu_barriers->contains(n), "duplicate entry in barrier list");
67 _iu_barriers->append(n);
68 }
69
70 void ShenandoahBarrierSetC2State::remove_iu_barrier(ShenandoahIUBarrierNode* n) {
71 _iu_barriers->remove_if_existing(n);
72 }
73
74 int ShenandoahBarrierSetC2State::load_reference_barriers_count() const {
75 return _load_reference_barriers->length();
76 }
77
78 ShenandoahLoadReferenceBarrierNode* ShenandoahBarrierSetC2State::load_reference_barrier(int idx) const {
79 return _load_reference_barriers->at(idx);
80 }
81
82 void ShenandoahBarrierSetC2State::add_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
83 assert(!_load_reference_barriers->contains(n), "duplicate entry in barrier list");
84 _load_reference_barriers->append(n);
85 }
86
87 void ShenandoahBarrierSetC2State::remove_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
88 if (_load_reference_barriers->contains(n)) {
89 _load_reference_barriers->remove(n);
90 }
91 }
92
93 Node* ShenandoahBarrierSetC2::shenandoah_iu_barrier(GraphKit* kit, Node* obj) const {
94 if (ShenandoahIUBarrier) {
95 return kit->gvn().transform(new ShenandoahIUBarrierNode(obj));
96 }
97 return obj;
98 }
99
100 #define __ kit->
101
102 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseValues* phase, Node* adr,
103 BasicType bt, uint adr_idx) const {
104 intptr_t offset = 0;
105 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
106 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
107
108 if (offset == Type::OffsetBot) {
109 return false; // cannot unalias unless there are precise offsets
110 }
111
112 if (alloc == nullptr) {
113 return false; // No allocation found
114 }
115
116 intptr_t size_in_bytes = type2aelembytes(bt);
117
118 Node* mem = __ memory(adr_idx); // start searching here...
119
120 for (int cnt = 0; cnt < 50; cnt++) {
121
122 if (mem->is_Store()) {
123
124 Node* st_adr = mem->in(MemNode::Address);
125 intptr_t st_offset = 0;
126 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
127
128 if (st_base == nullptr) {
129 break; // inscrutable pointer
130 }
131
132 // Break we have found a store with same base and offset as ours so break
133 if (st_base == base && st_offset == offset) {
134 break;
135 }
136
137 if (st_offset != offset && st_offset != Type::OffsetBot) {
138 const int MAX_STORE = BytesPerLong;
139 if (st_offset >= offset + size_in_bytes ||
140 st_offset <= offset - MAX_STORE ||
141 st_offset <= offset - mem->as_Store()->memory_size()) {
142 // Success: The offsets are provably independent.
143 // (You may ask, why not just test st_offset != offset and be done?
144 // The answer is that stores of different sizes can co-exist
145 // in the same sequence of RawMem effects. We sometimes initialize
146 // a whole 'tile' of array elements with a single jint or jlong.)
147 mem = mem->in(MemNode::Memory);
148 continue; // advance through independent store memory
149 }
150 }
151
152 if (st_base != base
153 && MemNode::detect_ptr_independence(base, alloc, st_base,
154 AllocateNode::Ideal_allocation(st_base, phase),
155 phase)) {
156 // Success: The bases are provably independent.
157 mem = mem->in(MemNode::Memory);
158 continue; // advance through independent store memory
159 }
160 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
161
162 InitializeNode* st_init = mem->in(0)->as_Initialize();
163 AllocateNode* st_alloc = st_init->allocation();
164
165 // Make sure that we are looking at the same allocation site.
166 // The alloc variable is guaranteed to not be null here from earlier check.
167 if (alloc == st_alloc) {
168 // Check that the initialization is storing null so that no previous store
169 // has been moved up and directly write a reference
170 Node* captured_store = st_init->find_captured_store(offset,
171 type2aelembytes(T_OBJECT),
172 phase);
173 if (captured_store == nullptr || captured_store == st_init->zero_memory()) {
174 return true;
175 }
176 }
177 }
178
179 // Unless there is an explicit 'continue', we must bail out here,
180 // because 'mem' is an inscrutable memory state (e.g., a call).
181 break;
182 }
183
184 return false;
185 }
186
187 #undef __
188 #define __ ideal.
189
190 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit,
191 bool do_load,
192 Node* obj,
193 Node* adr,
194 uint alias_idx,
195 Node* val,
196 const TypeOopPtr* val_type,
197 Node* pre_val,
198 BasicType bt) const {
199 // Some sanity checks
200 // Note: val is unused in this routine.
201
202 if (do_load) {
203 // We need to generate the load of the previous value
204 assert(adr != nullptr, "where are loading from?");
205 assert(pre_val == nullptr, "loaded already?");
206 assert(val_type != nullptr, "need a type");
207
208 if (ReduceInitialCardMarks
209 && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
210 return;
211 }
212
213 } else {
214 // In this case both val_type and alias_idx are unused.
215 assert(pre_val != nullptr, "must be loaded already");
216 // Nothing to be done if pre_val is null.
217 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
218 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
219 }
220 assert(bt == T_OBJECT, "or we shouldn't be here");
221
222 IdealKit ideal(kit, true);
223
224 Node* tls = __ thread(); // ThreadLocalStorage
225
226 Node* no_base = __ top();
227 Node* zero = __ ConI(0);
228 Node* zeroX = __ ConX(0);
229
230 float likely = PROB_LIKELY(0.999);
231 float unlikely = PROB_UNLIKELY(0.999);
232
233 // Offsets into the thread
234 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
235 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
236
237 // Now the actual pointers into the thread
238 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
239 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
240
241 // Now some of the values
242 Node* marking;
243 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())));
244 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
245 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
246 assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape");
247
248 // if (!marking)
249 __ if_then(marking, BoolTest::ne, zero, unlikely); {
250 BasicType index_bt = TypeX_X->basic_type();
251 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading Shenandoah SATBMarkQueue::_index with wrong size.");
252 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
253
254 if (do_load) {
255 // load original value
256 // alias_idx correct??
257 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
258 }
259
260 // if (pre_val != nullptr)
261 __ if_then(pre_val, BoolTest::ne, kit->null()); {
262 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
263
264 // is the queue for this thread full?
265 __ if_then(index, BoolTest::ne, zeroX, likely); {
266
267 // decrement the index
268 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
269
270 // Now get the buffer location we will log the previous value into and store it
271 Node *log_addr = __ AddP(no_base, buffer, next_index);
272 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
273 // update the index
274 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
275
276 } __ else_(); {
277
278 // logging buffer is full, call the runtime
279 const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type();
280 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls);
281 } __ end_if(); // (!index)
282 } __ end_if(); // (pre_val != nullptr)
283 } __ end_if(); // (!marking)
284
285 // Final sync IdealKit and GraphKit.
286 kit->final_sync(ideal);
287
288 if (ShenandoahSATBBarrier && adr != nullptr) {
289 Node* c = kit->control();
290 Node* call = c->in(1)->in(1)->in(1)->in(0);
291 assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected");
292 call->add_req(adr);
293 }
294 }
295
296 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) {
297 return call->is_CallLeaf() &&
298 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry);
299 }
300
301 bool ShenandoahBarrierSetC2::is_shenandoah_clone_call(Node* call) {
302 return call->is_CallLeaf() &&
303 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier);
304 }
305
306 bool ShenandoahBarrierSetC2::is_shenandoah_lrb_call(Node* call) {
307 if (!call->is_CallLeaf()) {
308 return false;
309 }
310
311 address entry_point = call->as_CallLeaf()->entry_point();
312 return (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong)) ||
313 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow)) ||
314 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak)) ||
315 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow)) ||
316 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom)) ||
317 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow));
318 }
319
320 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseValues* phase, Node* n) {
321 if (n->Opcode() != Op_If) {
322 return false;
323 }
324
325 Node* bol = n->in(1);
326 assert(bol->is_Bool(), "");
327 Node* cmpx = bol->in(1);
328 if (bol->as_Bool()->_test._test == BoolTest::ne &&
329 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
330 is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
331 cmpx->in(1)->in(2)->is_Con() &&
332 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
333 return true;
334 }
335
336 return false;
337 }
338
339 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
340 if (!n->is_Load()) return false;
341 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
342 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
343 && n->in(2)->in(3)->is_Con()
344 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
345 }
346
347 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
348 bool do_load,
349 Node* obj,
350 Node* adr,
351 uint alias_idx,
352 Node* val,
353 const TypeOopPtr* val_type,
354 Node* pre_val,
355 BasicType bt) const {
356 if (ShenandoahSATBBarrier) {
357 IdealKit ideal(kit);
358 kit->sync_kit(ideal);
359
360 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
361
362 ideal.sync_kit(kit);
363 kit->final_sync(ideal);
364 }
365 }
366
367 // Helper that guards and inserts a pre-barrier.
368 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
369 Node* pre_val, bool need_mem_bar) const {
370 // We could be accessing the referent field of a reference object. If so, when Shenandoah
371 // is enabled, we need to log the value in the referent field in an SATB buffer.
372 // This routine performs some compile time filters and generates suitable
373 // runtime filters that guard the pre-barrier code.
374 // Also add memory barrier for non volatile load from the referent field
375 // to prevent commoning of loads across safepoint.
376
377 // Some compile time checks.
378
379 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
380 const TypeX* otype = offset->find_intptr_t_type();
381 if (otype != nullptr && otype->is_con() &&
382 otype->get_con() != java_lang_ref_Reference::referent_offset()) {
383 // Constant offset but not the reference_offset so just return
384 return;
385 }
386
387 // We only need to generate the runtime guards for instances.
388 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
389 if (btype != nullptr) {
390 if (btype->isa_aryptr()) {
391 // Array type so nothing to do
392 return;
393 }
394
395 const TypeInstPtr* itype = btype->isa_instptr();
396 if (itype != nullptr) {
397 // Can the klass of base_oop be statically determined to be
398 // _not_ a sub-class of Reference and _not_ Object?
399 ciKlass* klass = itype->instance_klass();
400 if (klass->is_loaded() &&
401 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
402 !kit->env()->Object_klass()->is_subtype_of(klass)) {
403 return;
404 }
405 }
406 }
407
408 // The compile time filters did not reject base_oop/offset so
409 // we need to generate the following runtime filters
410 //
411 // if (offset == java_lang_ref_Reference::_reference_offset) {
412 // if (instance_of(base, java.lang.ref.Reference)) {
413 // pre_barrier(_, pre_val, ...);
414 // }
415 // }
416
417 float likely = PROB_LIKELY( 0.999);
418 float unlikely = PROB_UNLIKELY(0.999);
419
420 IdealKit ideal(kit);
421
422 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset());
423
424 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
425 // Update graphKit memory and control from IdealKit.
426 kit->sync_kit(ideal);
427
428 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
429 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
430
431 // Update IdealKit memory and control from graphKit.
432 __ sync_kit(kit);
433
434 Node* one = __ ConI(1);
435 // is_instof == 0 if base_oop == nullptr
436 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
437
438 // Update graphKit from IdeakKit.
439 kit->sync_kit(ideal);
440
441 // Use the pre-barrier to record the value in the referent field
442 satb_write_barrier_pre(kit, false /* do_load */,
443 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */,
444 pre_val /* pre_val */,
445 T_OBJECT);
446 if (need_mem_bar) {
447 // Add memory barrier to prevent commoning reads from this field
448 // across safepoint since GC can change its value.
449 kit->insert_mem_bar(Op_MemBarCPUOrder);
450 }
451 // Update IdealKit from graphKit.
452 __ sync_kit(kit);
453
454 } __ end_if(); // _ref_type != ref_none
455 } __ end_if(); // offset == referent_offset
456
457 // Final sync IdealKit and GraphKit.
458 kit->final_sync(ideal);
459 }
460
461 Node* ShenandoahBarrierSetC2::byte_map_base_node(GraphKit* kit) const {
462 BarrierSet* bs = BarrierSet::barrier_set();
463 ShenandoahBarrierSet* ctbs = barrier_set_cast<ShenandoahBarrierSet>(bs);
464 CardTable::CardValue* card_table_base = ctbs->card_table()->byte_map_base();
465 if (card_table_base != nullptr) {
466 return kit->makecon(TypeRawPtr::make((address)card_table_base));
467 } else {
468 return kit->null();
469 }
470 }
471
472 void ShenandoahBarrierSetC2::post_barrier(GraphKit* kit,
473 Node* ctl,
474 Node* oop_store,
475 Node* obj,
476 Node* adr,
477 uint adr_idx,
478 Node* val,
479 BasicType bt,
480 bool use_precise) const {
481 assert(ShenandoahCardBarrier, "Should have been checked by caller");
482
483 // No store check needed if we're storing a null.
484 if (val != nullptr && val->is_Con()) {
485 // must be either an oop or NULL
486 const Type* t = val->bottom_type();
487 if (t == TypePtr::NULL_PTR || t == Type::TOP)
488 return;
489 }
490
491 if (ReduceInitialCardMarks && obj == kit->just_allocated_object(kit->control())) {
492 // We can skip marks on a freshly-allocated object in Eden.
493 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
494 // That routine informs GC to take appropriate compensating steps,
495 // upon a slow-path allocation, so as to make this card-mark
496 // elision safe.
497 return;
498 }
499
500 if (!use_precise) {
501 // All card marks for a (non-array) instance are in one place:
502 adr = obj;
503 }
504 // (Else it's an array (or unknown), and we want more precise card marks.)
505 assert(adr != nullptr, "");
506
507 IdealKit ideal(kit, true);
508
509 // Convert the pointer to an int prior to doing math on it
510 Node* cast = __ CastPX(__ ctrl(), adr);
511
512 // Divide by card size
513 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift()) );
514
515 // Combine card table base and card offset
516 Node* card_adr = __ AddP(__ top(), byte_map_base_node(kit), card_offset );
517
518 // Get the alias_index for raw card-mark memory
519 int adr_type = Compile::AliasIdxRaw;
520 Node* zero = __ ConI(0); // Dirty card value
521
522 if (UseCondCardMark) {
523 // The classic GC reference write barrier is typically implemented
524 // as a store into the global card mark table. Unfortunately
525 // unconditional stores can result in false sharing and excessive
526 // coherence traffic as well as false transactional aborts.
527 // UseCondCardMark enables MP "polite" conditional card mark
528 // stores. In theory we could relax the load from ctrl() to
529 // no_ctrl, but that doesn't buy much latitude.
530 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, T_BYTE, adr_type);
531 __ if_then(card_val, BoolTest::ne, zero);
532 }
533
534 // Smash zero into card
535 __ store(__ ctrl(), card_adr, zero, T_BYTE, adr_type, MemNode::unordered);
536
537 if (UseCondCardMark) {
538 __ end_if();
539 }
540
541 // Final sync IdealKit and GraphKit.
542 kit->final_sync(ideal);
543 }
544
545 #undef __
546
547 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() {
548 const Type **fields = TypeTuple::fields(2);
549 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
550 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
551 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
552
553 // create result type (range)
554 fields = TypeTuple::fields(0);
555 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
556
557 return TypeFunc::make(domain, range);
558 }
559
560 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() {
561 const Type **fields = TypeTuple::fields(1);
562 fields[TypeFunc::Parms+0] = TypeOopPtr::NOTNULL; // src oop
563 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
564
565 // create result type (range)
566 fields = TypeTuple::fields(0);
567 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
568
569 return TypeFunc::make(domain, range);
570 }
571
572 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type() {
573 const Type **fields = TypeTuple::fields(2);
574 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM; // original field value
575 fields[TypeFunc::Parms+1] = TypeRawPtr::BOTTOM; // original load address
576
577 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
578
579 // create result type (range)
580 fields = TypeTuple::fields(1);
581 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM;
582 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
583
584 return TypeFunc::make(domain, range);
585 }
586
587 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
588 DecoratorSet decorators = access.decorators();
589
590 const TypePtr* adr_type = access.addr().type();
591 Node* adr = access.addr().node();
592
593 if (!access.is_oop()) {
594 return BarrierSetC2::store_at_resolved(access, val);
595 }
596
597 if (access.is_parse_access()) {
598 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
599 GraphKit* kit = parse_access.kit();
600
601 uint adr_idx = kit->C->get_alias_index(adr_type);
602 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
603 Node* value = val.node();
604 value = shenandoah_iu_barrier(kit, value);
605 val.set_node(value);
606 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
607 static_cast<const TypeOopPtr*>(val.type()), nullptr /* pre_val */, access.type());
608
609 Node* result = BarrierSetC2::store_at_resolved(access, val);
610
611 if (ShenandoahCardBarrier) {
612 const bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
613 const bool is_array = (decorators & IS_ARRAY) != 0;
614 const bool use_precise = is_array || anonymous;
615 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
616 adr, adr_idx, val.node(), access.type(), use_precise);
617 }
618 return result;
619 } else {
620 assert(access.is_opt_access(), "only for optimization passes");
621 assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code");
622 C2OptAccess& opt_access = static_cast<C2OptAccess&>(access);
623 PhaseGVN& gvn = opt_access.gvn();
624
625 if (ShenandoahIUBarrier) {
626 Node* enqueue = gvn.transform(new ShenandoahIUBarrierNode(val.node()));
627 val.set_node(enqueue);
628 }
629 return BarrierSetC2::store_at_resolved(access, val);
630 }
631 }
632
633 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
634 // 1: non-reference load, no additional barrier is needed
635 if (!access.is_oop()) {
636 return BarrierSetC2::load_at_resolved(access, val_type);
637 }
638
639 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
640 DecoratorSet decorators = access.decorators();
641 BasicType type = access.type();
642
643 // 2: apply LRB if needed
644 if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
645 load = new ShenandoahLoadReferenceBarrierNode(nullptr, load, decorators);
646 if (access.is_parse_access()) {
647 load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load);
648 } else {
649 load = static_cast<C2OptAccess &>(access).gvn().transform(load);
650 }
651 }
652
653 // 3: apply keep-alive barrier for java.lang.ref.Reference if needed
654 if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
655 Node* top = Compile::current()->top();
656 Node* adr = access.addr().node();
657 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
658 Node* obj = access.base();
659
660 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
661 bool on_weak_ref = (decorators & (ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF)) != 0;
662 bool keep_alive = (decorators & AS_NO_KEEPALIVE) == 0;
663
664 // If we are reading the value of the referent field of a Reference
665 // object (either by using Unsafe directly or through reflection)
666 // then, if SATB is enabled, we need to record the referent in an
667 // SATB log buffer using the pre-barrier mechanism.
668 // Also we need to add memory barrier to prevent commoning reads
669 // from this field across safepoint since GC can change its value.
670 if (!on_weak_ref || (unknown && (offset == top || obj == top)) || !keep_alive) {
671 return load;
672 }
673
674 assert(access.is_parse_access(), "entry not supported at optimization time");
675 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
676 GraphKit* kit = parse_access.kit();
677 bool mismatched = (decorators & C2_MISMATCHED) != 0;
678 bool is_unordered = (decorators & MO_UNORDERED) != 0;
679 bool in_native = (decorators & IN_NATIVE) != 0;
680 bool need_cpu_mem_bar = !is_unordered || mismatched || in_native;
681
682 if (on_weak_ref) {
683 // Use the pre-barrier to record the value in the referent field
684 satb_write_barrier_pre(kit, false /* do_load */,
685 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */,
686 load /* pre_val */, T_OBJECT);
687 // Add memory barrier to prevent commoning reads from this field
688 // across safepoint since GC can change its value.
689 kit->insert_mem_bar(Op_MemBarCPUOrder);
690 } else if (unknown) {
691 // We do not require a mem bar inside pre_barrier if need_mem_bar
692 // is set: the barriers would be emitted by us.
693 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
694 }
695 }
696
697 return load;
698 }
699
700 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
701 Node* new_val, const Type* value_type) const {
702 GraphKit* kit = access.kit();
703 if (access.is_oop()) {
704 new_val = shenandoah_iu_barrier(kit, new_val);
705 shenandoah_write_barrier_pre(kit, false /* do_load */,
706 nullptr, nullptr, max_juint, nullptr, nullptr,
707 expected_val /* pre_val */, T_OBJECT);
708
709 MemNode::MemOrd mo = access.mem_node_mo();
710 Node* mem = access.memory();
711 Node* adr = access.addr().node();
712 const TypePtr* adr_type = access.addr().type();
713 Node* load_store = nullptr;
714
715 #ifdef _LP64
716 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
717 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
718 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
719 if (ShenandoahCASBarrier) {
720 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
721 } else {
722 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
723 }
724 } else
725 #endif
726 {
727 if (ShenandoahCASBarrier) {
728 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
729 } else {
730 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
731 }
732 }
733
734 access.set_raw_access(load_store);
735 pin_atomic_op(access);
736
737 #ifdef _LP64
738 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
739 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
740 }
741 #endif
742 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, load_store, access.decorators()));
743 if (ShenandoahCardBarrier) {
744 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
745 access.addr().node(), access.alias_idx(), new_val, T_OBJECT, true);
746 }
747 return load_store;
748 }
749 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
750 }
751
752 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
753 Node* new_val, const Type* value_type) const {
754 GraphKit* kit = access.kit();
755 if (access.is_oop()) {
756 new_val = shenandoah_iu_barrier(kit, new_val);
757 shenandoah_write_barrier_pre(kit, false /* do_load */,
758 nullptr, nullptr, max_juint, nullptr, nullptr,
759 expected_val /* pre_val */, T_OBJECT);
760 DecoratorSet decorators = access.decorators();
761 MemNode::MemOrd mo = access.mem_node_mo();
762 Node* mem = access.memory();
763 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
764 Node* load_store = nullptr;
765 Node* adr = access.addr().node();
766 #ifdef _LP64
767 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
768 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
769 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
770 if (ShenandoahCASBarrier) {
771 if (is_weak_cas) {
772 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
773 } else {
774 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
775 }
776 } else {
777 if (is_weak_cas) {
778 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
779 } else {
780 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
781 }
782 }
783 } else
784 #endif
785 {
786 if (ShenandoahCASBarrier) {
787 if (is_weak_cas) {
788 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
789 } else {
790 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
791 }
792 } else {
793 if (is_weak_cas) {
794 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
795 } else {
796 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
797 }
798 }
799 }
800 access.set_raw_access(load_store);
801 pin_atomic_op(access);
802 if (ShenandoahCardBarrier) {
803 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
804 access.addr().node(), access.alias_idx(), new_val, T_OBJECT, true);
805 }
806 return load_store;
807 }
808 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
809 }
810
811 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
812 GraphKit* kit = access.kit();
813 if (access.is_oop()) {
814 val = shenandoah_iu_barrier(kit, val);
815 }
816 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
817 if (access.is_oop()) {
818 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, result, access.decorators()));
819 shenandoah_write_barrier_pre(kit, false /* do_load */,
820 nullptr, nullptr, max_juint, nullptr, nullptr,
821 result /* pre_val */, T_OBJECT);
822 if (ShenandoahCardBarrier) {
823 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
824 access.addr().node(), access.alias_idx(), val, T_OBJECT, true);
825 }
826 }
827 return result;
828 }
829
830
831 bool ShenandoahBarrierSetC2::is_gc_pre_barrier_node(Node* node) const {
832 return is_shenandoah_wb_pre_call(node);
833 }
834
835 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
836 return (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) || (node->Opcode() == Op_ShenandoahIUBarrier) ||
837 is_shenandoah_lrb_call(node) ||
838 is_shenandoah_wb_pre_call(node) ||
839 is_shenandoah_clone_call(node);
840 }
841
842 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
843 if (c == nullptr) {
844 return c;
845 }
846 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
847 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
848 }
849 if (c->Opcode() == Op_ShenandoahIUBarrier) {
850 c = c->in(1);
851 }
852 return c;
853 }
854
855 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
856 return !ShenandoahBarrierC2Support::expand(C, igvn);
857 }
858
859 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
860 if (mode == LoopOptsShenandoahExpand) {
861 assert(UseShenandoahGC, "only for shenandoah");
862 ShenandoahBarrierC2Support::pin_and_expand(phase);
863 return true;
864 } else if (mode == LoopOptsShenandoahPostExpand) {
865 assert(UseShenandoahGC, "only for shenandoah");
866 visited.clear();
867 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
868 return true;
869 }
870 return false;
871 }
872
873 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, bool is_clone_instance, ArrayCopyPhase phase) const {
874 bool is_oop = is_reference_type(type);
875 if (!is_oop) {
876 return false;
877 }
878 if (ShenandoahSATBBarrier && tightly_coupled_alloc) {
879 if (phase == Optimization) {
880 return false;
881 }
882 return !is_clone;
883 }
884 if (phase == Optimization) {
885 return !ShenandoahIUBarrier;
886 }
887 return true;
888 }
889
890 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
891 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
892 if (src_type->isa_instptr() != nullptr) {
893 ciInstanceKlass* ik = src_type->is_instptr()->instance_klass();
894 if ((src_type->klass_is_exact() || !ik->has_subklass()) && !ik->has_injected_fields()) {
895 if (ik->has_object_fields()) {
896 return true;
897 } else {
898 if (!src_type->klass_is_exact()) {
899 Compile::current()->dependencies()->assert_leaf_type(ik);
900 }
901 }
902 } else {
903 return true;
904 }
905 } else if (src_type->isa_aryptr()) {
906 BasicType src_elem = src_type->isa_aryptr()->elem()->array_element_basic_type();
907 if (is_reference_type(src_elem, true)) {
908 return true;
909 }
910 } else {
911 return true;
912 }
913 return false;
914 }
915
916 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
917 Node* ctrl = ac->in(TypeFunc::Control);
918 Node* mem = ac->in(TypeFunc::Memory);
919 Node* src_base = ac->in(ArrayCopyNode::Src);
920 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
921 Node* dest_base = ac->in(ArrayCopyNode::Dest);
922 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
923 Node* length = ac->in(ArrayCopyNode::Length);
924
925 Node* src = phase->basic_plus_adr(src_base, src_offset);
926 Node* dest = phase->basic_plus_adr(dest_base, dest_offset);
927
928 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
929 // Check if heap is has forwarded objects. If it does, we need to call into the special
930 // routine that would fix up source references before we can continue.
931
932 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
933 Node* region = new RegionNode(PATH_LIMIT);
934 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM);
935
936 Node* thread = phase->transform_later(new ThreadLocalNode());
937 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
938 Node* gc_state_addr = phase->transform_later(new AddPNode(phase->C->top(), thread, offset));
939
940 uint gc_state_idx = Compile::AliasIdxRaw;
941 const TypePtr* gc_state_adr_type = nullptr; // debug-mode-only argument
942 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
943
944 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered));
945 int flags = ShenandoahHeap::HAS_FORWARDED;
946 if (ShenandoahIUBarrier) {
947 flags |= ShenandoahHeap::MARKING;
948 }
949 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(flags)));
950 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT)));
951 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne));
952
953 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If();
954 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff));
955 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff));
956
957 // Heap is stable, no need to do anything additional
958 region->init_req(_heap_stable, stable_ctrl);
959 mem_phi->init_req(_heap_stable, mem);
960
961 // Heap is unstable, call into clone barrier stub
962 Node* call = phase->make_leaf_call(unstable_ctrl, mem,
963 ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(),
964 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier),
965 "shenandoah_clone",
966 TypeRawPtr::BOTTOM,
967 src_base);
968 call = phase->transform_later(call);
969
970 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control));
971 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory));
972 region->init_req(_heap_unstable, ctrl);
973 mem_phi->init_req(_heap_unstable, mem);
974
975 // Wire up the actual arraycopy stub now
976 ctrl = phase->transform_later(region);
977 mem = phase->transform_later(mem_phi);
978
979 const char* name = "arraycopy";
980 call = phase->make_leaf_call(ctrl, mem,
981 OptoRuntime::fast_arraycopy_Type(),
982 phase->basictype2arraycopy(T_LONG, nullptr, nullptr, true, name, true),
983 name, TypeRawPtr::BOTTOM,
984 src, dest, length
985 LP64_ONLY(COMMA phase->top()));
986 call = phase->transform_later(call);
987
988 // Hook up the whole thing into the graph
989 phase->igvn().replace_node(ac, call);
990 } else {
991 BarrierSetC2::clone_at_expansion(phase, ac);
992 }
993 }
994
995
996 // Support for macro expanded GC barriers
997 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
998 if (node->Opcode() == Op_ShenandoahIUBarrier) {
999 state()->add_iu_barrier((ShenandoahIUBarrierNode*) node);
1000 }
1001 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
1002 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
1003 }
1004 }
1005
1006 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
1007 if (node->Opcode() == Op_ShenandoahIUBarrier) {
1008 state()->remove_iu_barrier((ShenandoahIUBarrierNode*) node);
1009 }
1010 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
1011 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
1012 }
1013 }
1014
1015 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const {
1016 if (is_shenandoah_wb_pre_call(node)) {
1017 shenandoah_eliminate_wb_pre(node, ¯o->igvn());
1018 }
1019 if (ShenandoahCardBarrier && node->Opcode() == Op_CastP2X) {
1020 Node* shift = node->unique_out();
1021 Node* addp = shift->unique_out();
1022 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
1023 Node* mem = addp->last_out(j);
1024 if (UseCondCardMark && mem->is_Load()) {
1025 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
1026 // The load is checking if the card has been written so
1027 // replace it with zero to fold the test.
1028 macro->replace_node(mem, macro->intcon(0));
1029 continue;
1030 }
1031 assert(mem->is_Store(), "store required");
1032 macro->replace_node(mem, mem->in(MemNode::Memory));
1033 }
1034 }
1035 }
1036
1037 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
1038 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
1039 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
1040 c = c->unique_ctrl_out();
1041 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
1042 c = c->unique_ctrl_out();
1043 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
1044 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
1045 assert(iff->is_If(), "expect test");
1046 if (!is_shenandoah_marking_if(igvn, iff)) {
1047 c = c->unique_ctrl_out();
1048 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
1049 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
1050 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
1051 }
1052 Node* cmpx = iff->in(1)->in(1);
1053 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
1054 igvn->rehash_node_delayed(call);
1055 call->del_req(call->req()-1);
1056 }
1057
1058 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
1059 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
1060 igvn->add_users_to_worklist(node);
1061 }
1062 }
1063
1064 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
1065 for (uint i = 0; i < useful.size(); i++) {
1066 Node* n = useful.at(i);
1067 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
1068 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1069 C->record_for_igvn(n->fast_out(i));
1070 }
1071 }
1072 }
1073 for (int i = state()->iu_barriers_count() - 1; i >= 0; i--) {
1074 ShenandoahIUBarrierNode* n = state()->iu_barrier(i);
1075 if (!useful.member(n)) {
1076 state()->remove_iu_barrier(n);
1077 }
1078 }
1079 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
1080 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
1081 if (!useful.member(n)) {
1082 state()->remove_load_reference_barrier(n);
1083 }
1084 }
1085 }
1086
1087 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
1088 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
1089 }
1090
1091 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
1092 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
1093 }
1094
1095 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
1096 // expanded later, then now is the time to do so.
1097 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
1098
1099 #ifdef ASSERT
1100 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
1101 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) {
1102 ShenandoahBarrierC2Support::verify(Compile::current()->root());
1103 } else if (phase == BarrierSetC2::BeforeCodeGen) {
1104 // Verify Shenandoah pre-barriers
1105 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset());
1106
1107 Unique_Node_List visited;
1108 Node_List worklist;
1109 // We're going to walk control flow backwards starting from the Root
1110 worklist.push(compile->root());
1111 while (worklist.size() > 0) {
1112 Node *x = worklist.pop();
1113 if (x == nullptr || x == compile->top()) continue;
1114 if (visited.member(x)) {
1115 continue;
1116 } else {
1117 visited.push(x);
1118 }
1119
1120 if (x->is_Region()) {
1121 for (uint i = 1; i < x->req(); i++) {
1122 worklist.push(x->in(i));
1123 }
1124 } else {
1125 worklist.push(x->in(0));
1126 // We are looking for the pattern:
1127 // /->ThreadLocal
1128 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
1129 // \->ConI(0)
1130 // We want to verify that the If and the LoadB have the same control
1131 // See GraphKit::g1_write_barrier_pre()
1132 if (x->is_If()) {
1133 IfNode *iff = x->as_If();
1134 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
1135 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
1136 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
1137 && cmp->in(1)->is_Load()) {
1138 LoadNode *load = cmp->in(1)->as_Load();
1139 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
1140 && load->in(2)->in(3)->is_Con()
1141 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
1142
1143 Node *if_ctrl = iff->in(0);
1144 Node *load_ctrl = load->in(0);
1145
1146 if (if_ctrl != load_ctrl) {
1147 // Skip possible CProj->NeverBranch in infinite loops
1148 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
1149 && if_ctrl->in(0)->is_NeverBranch()) {
1150 if_ctrl = if_ctrl->in(0)->in(0);
1151 }
1152 }
1153 assert(load_ctrl != nullptr && if_ctrl == load_ctrl, "controls must match");
1154 }
1155 }
1156 }
1157 }
1158 }
1159 }
1160 }
1161 }
1162 #endif
1163
1164 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
1165 if (is_shenandoah_wb_pre_call(n)) {
1166 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1167 if (n->req() > cnt) {
1168 Node* addp = n->in(cnt);
1169 if (has_only_shenandoah_wb_pre_uses(addp)) {
1170 n->del_req(cnt);
1171 if (can_reshape) {
1172 phase->is_IterGVN()->_worklist.push(addp);
1173 }
1174 return n;
1175 }
1176 }
1177 }
1178 if (n->Opcode() == Op_CmpP) {
1179 Node* in1 = n->in(1);
1180 Node* in2 = n->in(2);
1181
1182 // If one input is null, then step over the strong LRB barriers on the other input
1183 if (in1->bottom_type() == TypePtr::NULL_PTR &&
1184 !((in2->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1185 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in2)->decorators()))) {
1186 in2 = step_over_gc_barrier(in2);
1187 }
1188 if (in2->bottom_type() == TypePtr::NULL_PTR &&
1189 !((in1->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1190 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in1)->decorators()))) {
1191 in1 = step_over_gc_barrier(in1);
1192 }
1193
1194 if (in1 != n->in(1)) {
1195 n->set_req_X(1, in1, phase);
1196 assert(in2 == n->in(2), "only one change");
1197 return n;
1198 }
1199 if (in2 != n->in(2)) {
1200 n->set_req_X(2, in2, phase);
1201 return n;
1202 }
1203 } else if (can_reshape &&
1204 n->Opcode() == Op_If &&
1205 ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1206 n->in(0) != nullptr &&
1207 n->outcnt() == 2) {
1208 Node* dom = n->in(0);
1209 Node* prev_dom = n;
1210 int op = n->Opcode();
1211 int dist = 16;
1212 // Search up the dominator tree for another heap stable test
1213 while (dom->Opcode() != op || // Not same opcode?
1214 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1?
1215 prev_dom->in(0) != dom) { // One path of test does not dominate?
1216 if (dist < 0) return nullptr;
1217
1218 dist--;
1219 prev_dom = dom;
1220 dom = IfNode::up_one_dom(dom);
1221 if (!dom) return nullptr;
1222 }
1223
1224 // Check that we did not follow a loop back to ourselves
1225 if (n == dom) {
1226 return nullptr;
1227 }
1228
1229 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN(), false);
1230 }
1231
1232 return nullptr;
1233 }
1234
1235 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1236 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1237 Node* u = n->fast_out(i);
1238 if (!is_shenandoah_wb_pre_call(u)) {
1239 return false;
1240 }
1241 }
1242 return n->outcnt() > 0;
1243 }
1244
1245 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode, Unique_Node_List& dead_nodes) const {
1246 switch (opcode) {
1247 case Op_CallLeaf:
1248 case Op_CallLeafNoFP: {
1249 assert (n->is_Call(), "");
1250 CallNode *call = n->as_Call();
1251 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1252 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1253 if (call->req() > cnt) {
1254 assert(call->req() == cnt + 1, "only one extra input");
1255 Node *addp = call->in(cnt);
1256 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1257 call->del_req(cnt);
1258 }
1259 }
1260 return false;
1261 }
1262 case Op_ShenandoahCompareAndSwapP:
1263 case Op_ShenandoahCompareAndSwapN:
1264 case Op_ShenandoahWeakCompareAndSwapN:
1265 case Op_ShenandoahWeakCompareAndSwapP:
1266 case Op_ShenandoahCompareAndExchangeP:
1267 case Op_ShenandoahCompareAndExchangeN:
1268 return true;
1269 case Op_ShenandoahLoadReferenceBarrier:
1270 assert(false, "should have been expanded already");
1271 return true;
1272 default:
1273 return false;
1274 }
1275 }
1276
1277 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1278 switch (opcode) {
1279 case Op_ShenandoahCompareAndExchangeP:
1280 case Op_ShenandoahCompareAndExchangeN:
1281 conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1282 // fallthrough
1283 case Op_ShenandoahWeakCompareAndSwapP:
1284 case Op_ShenandoahWeakCompareAndSwapN:
1285 case Op_ShenandoahCompareAndSwapP:
1286 case Op_ShenandoahCompareAndSwapN:
1287 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1288 return true;
1289 case Op_StoreP: {
1290 Node* adr = n->in(MemNode::Address);
1291 const Type* adr_type = gvn->type(adr);
1292 // Pointer stores in Shenandoah barriers looks like unsafe access.
1293 // Ignore such stores to be able scalar replace non-escaping
1294 // allocations.
1295 if (adr_type->isa_rawptr() && adr->is_AddP()) {
1296 Node* base = conn_graph->get_addp_base(adr);
1297 if (base->Opcode() == Op_LoadP &&
1298 base->in(MemNode::Address)->is_AddP()) {
1299 adr = base->in(MemNode::Address);
1300 Node* tls = conn_graph->get_addp_base(adr);
1301 if (tls->Opcode() == Op_ThreadLocal) {
1302 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1303 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1304 if (offs == buf_offset) {
1305 return true; // Pre barrier previous oop value store.
1306 }
1307 }
1308 }
1309 }
1310 return false;
1311 }
1312 case Op_ShenandoahIUBarrier:
1313 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1314 break;
1315 case Op_ShenandoahLoadReferenceBarrier:
1316 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist);
1317 return true;
1318 default:
1319 // Nothing
1320 break;
1321 }
1322 return false;
1323 }
1324
1325 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1326 switch (opcode) {
1327 case Op_ShenandoahCompareAndExchangeP:
1328 case Op_ShenandoahCompareAndExchangeN: {
1329 Node *adr = n->in(MemNode::Address);
1330 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, nullptr);
1331 // fallthrough
1332 }
1333 case Op_ShenandoahCompareAndSwapP:
1334 case Op_ShenandoahCompareAndSwapN:
1335 case Op_ShenandoahWeakCompareAndSwapP:
1336 case Op_ShenandoahWeakCompareAndSwapN:
1337 return conn_graph->add_final_edges_unsafe_access(n, opcode);
1338 case Op_ShenandoahIUBarrier:
1339 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), nullptr);
1340 return true;
1341 case Op_ShenandoahLoadReferenceBarrier:
1342 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), nullptr);
1343 return true;
1344 default:
1345 // Nothing
1346 break;
1347 }
1348 return false;
1349 }
1350
1351 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1352 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1353 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1354
1355 }
1356
1357 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1358 switch (opcode) {
1359 case Op_ShenandoahCompareAndExchangeP:
1360 case Op_ShenandoahCompareAndExchangeN:
1361 case Op_ShenandoahWeakCompareAndSwapP:
1362 case Op_ShenandoahWeakCompareAndSwapN:
1363 case Op_ShenandoahCompareAndSwapP:
1364 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree
1365 Node* newval = n->in(MemNode::ValueIn);
1366 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1367 Node* pair = new BinaryNode(oldval, newval);
1368 n->set_req(MemNode::ValueIn,pair);
1369 n->del_req(LoadStoreConditionalNode::ExpectedIn);
1370 return true;
1371 }
1372 default:
1373 break;
1374 }
1375 return false;
1376 }
1377
1378 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1379 return xop == Op_ShenandoahCompareAndExchangeP ||
1380 xop == Op_ShenandoahCompareAndExchangeN ||
1381 xop == Op_ShenandoahWeakCompareAndSwapP ||
1382 xop == Op_ShenandoahWeakCompareAndSwapN ||
1383 xop == Op_ShenandoahCompareAndSwapN ||
1384 xop == Op_ShenandoahCompareAndSwapP;
1385 }