1 /*
2 * Copyright (c) 2018, 2023, Red Hat, Inc. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/javaClasses.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shenandoah/shenandoahBarrierSet.hpp"
29 #include "gc/shenandoah/shenandoahForwarding.hpp"
30 #include "gc/shenandoah/shenandoahHeap.hpp"
31 #include "gc/shenandoah/shenandoahRuntime.hpp"
32 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
33 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
34 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
35 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
36 #include "opto/arraycopynode.hpp"
37 #include "opto/escape.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/idealKit.hpp"
40 #include "opto/macro.hpp"
41 #include "opto/movenode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/rootnode.hpp"
44 #include "opto/runtime.hpp"
45
46 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() {
47 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2());
48 }
49
50 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena)
51 : _iu_barriers(new (comp_arena) GrowableArray<ShenandoahIUBarrierNode*>(comp_arena, 8, 0, nullptr)),
52 _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8, 0, nullptr)) {
53 }
54
55 int ShenandoahBarrierSetC2State::iu_barriers_count() const {
56 return _iu_barriers->length();
57 }
58
59 ShenandoahIUBarrierNode* ShenandoahBarrierSetC2State::iu_barrier(int idx) const {
60 return _iu_barriers->at(idx);
61 }
62
63 void ShenandoahBarrierSetC2State::add_iu_barrier(ShenandoahIUBarrierNode* n) {
64 assert(!_iu_barriers->contains(n), "duplicate entry in barrier list");
65 _iu_barriers->append(n);
66 }
67
68 void ShenandoahBarrierSetC2State::remove_iu_barrier(ShenandoahIUBarrierNode* n) {
69 _iu_barriers->remove_if_existing(n);
70 }
71
72 int ShenandoahBarrierSetC2State::load_reference_barriers_count() const {
73 return _load_reference_barriers->length();
74 }
75
76 ShenandoahLoadReferenceBarrierNode* ShenandoahBarrierSetC2State::load_reference_barrier(int idx) const {
77 return _load_reference_barriers->at(idx);
78 }
79
80 void ShenandoahBarrierSetC2State::add_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
81 assert(!_load_reference_barriers->contains(n), "duplicate entry in barrier list");
82 _load_reference_barriers->append(n);
83 }
84
85 void ShenandoahBarrierSetC2State::remove_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
86 if (_load_reference_barriers->contains(n)) {
87 _load_reference_barriers->remove(n);
88 }
89 }
90
91 Node* ShenandoahBarrierSetC2::shenandoah_iu_barrier(GraphKit* kit, Node* obj) const {
92 if (ShenandoahIUBarrier) {
93 return kit->gvn().transform(new ShenandoahIUBarrierNode(obj));
94 }
95 return obj;
96 }
97
98 #define __ kit->
99
100 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseValues* phase, Node* adr,
101 BasicType bt, uint adr_idx) const {
102 intptr_t offset = 0;
103 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
104 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
105
106 if (offset == Type::OffsetBot) {
107 return false; // cannot unalias unless there are precise offsets
108 }
109
110 if (alloc == nullptr) {
111 return false; // No allocation found
112 }
113
114 intptr_t size_in_bytes = type2aelembytes(bt);
115
116 Node* mem = __ memory(adr_idx); // start searching here...
117
118 for (int cnt = 0; cnt < 50; cnt++) {
119
120 if (mem->is_Store()) {
121
122 Node* st_adr = mem->in(MemNode::Address);
123 intptr_t st_offset = 0;
124 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
125
126 if (st_base == nullptr) {
127 break; // inscrutable pointer
128 }
129
130 // Break we have found a store with same base and offset as ours so break
131 if (st_base == base && st_offset == offset) {
132 break;
133 }
134
135 if (st_offset != offset && st_offset != Type::OffsetBot) {
136 const int MAX_STORE = BytesPerLong;
137 if (st_offset >= offset + size_in_bytes ||
138 st_offset <= offset - MAX_STORE ||
139 st_offset <= offset - mem->as_Store()->memory_size()) {
140 // Success: The offsets are provably independent.
141 // (You may ask, why not just test st_offset != offset and be done?
142 // The answer is that stores of different sizes can co-exist
143 // in the same sequence of RawMem effects. We sometimes initialize
144 // a whole 'tile' of array elements with a single jint or jlong.)
145 mem = mem->in(MemNode::Memory);
146 continue; // advance through independent store memory
147 }
148 }
149
150 if (st_base != base
151 && MemNode::detect_ptr_independence(base, alloc, st_base,
152 AllocateNode::Ideal_allocation(st_base),
153 phase)) {
154 // Success: The bases are provably independent.
155 mem = mem->in(MemNode::Memory);
156 continue; // advance through independent store memory
157 }
158 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
159
160 InitializeNode* st_init = mem->in(0)->as_Initialize();
161 AllocateNode* st_alloc = st_init->allocation();
162
163 // Make sure that we are looking at the same allocation site.
164 // The alloc variable is guaranteed to not be null here from earlier check.
165 if (alloc == st_alloc) {
166 // Check that the initialization is storing null so that no previous store
167 // has been moved up and directly write a reference
168 Node* captured_store = st_init->find_captured_store(offset,
169 type2aelembytes(T_OBJECT),
170 phase);
171 if (captured_store == nullptr || captured_store == st_init->zero_memory()) {
172 return true;
173 }
174 }
175 }
176
177 // Unless there is an explicit 'continue', we must bail out here,
178 // because 'mem' is an inscrutable memory state (e.g., a call).
179 break;
180 }
181
182 return false;
183 }
184
185 #undef __
186 #define __ ideal.
187
188 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit,
189 bool do_load,
190 Node* obj,
191 Node* adr,
192 uint alias_idx,
193 Node* val,
194 const TypeOopPtr* val_type,
195 Node* pre_val,
196 BasicType bt) const {
197 // Some sanity checks
198 // Note: val is unused in this routine.
199
200 if (do_load) {
201 // We need to generate the load of the previous value
202 assert(adr != nullptr, "where are loading from?");
203 assert(pre_val == nullptr, "loaded already?");
204 assert(val_type != nullptr, "need a type");
205
206 if (ReduceInitialCardMarks
207 && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
208 return;
209 }
210
211 } else {
212 // In this case both val_type and alias_idx are unused.
213 assert(pre_val != nullptr, "must be loaded already");
214 // Nothing to be done if pre_val is null.
215 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
216 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
217 }
218 assert(bt == T_OBJECT, "or we shouldn't be here");
219
220 IdealKit ideal(kit, true);
221
222 Node* tls = __ thread(); // ThreadLocalStorage
223
224 Node* no_base = __ top();
225 Node* zero = __ ConI(0);
226 Node* zeroX = __ ConX(0);
227
228 float likely = PROB_LIKELY(0.999);
229 float unlikely = PROB_UNLIKELY(0.999);
230
231 // Offsets into the thread
232 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
233 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
234
235 // Now the actual pointers into the thread
236 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
237 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
238
239 // Now some of the values
240 Node* marking;
241 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())));
242 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
243 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
244 assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape");
245
246 // if (!marking)
247 __ if_then(marking, BoolTest::ne, zero, unlikely); {
248 BasicType index_bt = TypeX_X->basic_type();
249 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading Shenandoah SATBMarkQueue::_index with wrong size.");
250 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
251
252 if (do_load) {
253 // load original value
254 // alias_idx correct??
255 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
256 }
257
258 // if (pre_val != nullptr)
259 __ if_then(pre_val, BoolTest::ne, kit->null()); {
260 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
261
262 // is the queue for this thread full?
263 __ if_then(index, BoolTest::ne, zeroX, likely); {
264
265 // decrement the index
266 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
267
268 // Now get the buffer location we will log the previous value into and store it
269 Node *log_addr = __ AddP(no_base, buffer, next_index);
270 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
271 // update the index
272 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
273
274 } __ else_(); {
275
276 // logging buffer is full, call the runtime
277 const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type();
278 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls);
279 } __ end_if(); // (!index)
280 } __ end_if(); // (pre_val != nullptr)
281 } __ end_if(); // (!marking)
282
283 // Final sync IdealKit and GraphKit.
284 kit->final_sync(ideal);
285
286 if (ShenandoahSATBBarrier && adr != nullptr) {
287 Node* c = kit->control();
288 Node* call = c->in(1)->in(1)->in(1)->in(0);
289 assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected");
290 call->add_req(adr);
291 }
292 }
293
294 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) {
295 return call->is_CallLeaf() &&
296 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry);
297 }
298
299 bool ShenandoahBarrierSetC2::is_shenandoah_lrb_call(Node* call) {
300 if (!call->is_CallLeaf()) {
301 return false;
302 }
303
304 address entry_point = call->as_CallLeaf()->entry_point();
305 return (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong)) ||
306 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow)) ||
307 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak)) ||
308 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow)) ||
309 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom));
310 }
311
312 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseValues* phase, Node* n) {
313 if (n->Opcode() != Op_If) {
314 return false;
315 }
316
317 Node* bol = n->in(1);
318 assert(bol->is_Bool(), "");
319 Node* cmpx = bol->in(1);
320 if (bol->as_Bool()->_test._test == BoolTest::ne &&
321 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
322 is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
323 cmpx->in(1)->in(2)->is_Con() &&
324 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
325 return true;
326 }
327
328 return false;
329 }
330
331 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
332 if (!n->is_Load()) return false;
333 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
334 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
335 && n->in(2)->in(3)->is_Con()
336 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
337 }
338
339 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
340 bool do_load,
341 Node* obj,
342 Node* adr,
343 uint alias_idx,
344 Node* val,
345 const TypeOopPtr* val_type,
346 Node* pre_val,
347 BasicType bt) const {
348 if (ShenandoahSATBBarrier) {
349 IdealKit ideal(kit);
350 kit->sync_kit(ideal);
351
352 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
353
354 ideal.sync_kit(kit);
355 kit->final_sync(ideal);
356 }
357 }
358
359 // Helper that guards and inserts a pre-barrier.
360 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
361 Node* pre_val, bool need_mem_bar) const {
362 // We could be accessing the referent field of a reference object. If so, when Shenandoah
363 // is enabled, we need to log the value in the referent field in an SATB buffer.
364 // This routine performs some compile time filters and generates suitable
365 // runtime filters that guard the pre-barrier code.
366 // Also add memory barrier for non volatile load from the referent field
367 // to prevent commoning of loads across safepoint.
368
369 // Some compile time checks.
370
371 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
372 const TypeX* otype = offset->find_intptr_t_type();
373 if (otype != nullptr && otype->is_con() &&
374 otype->get_con() != java_lang_ref_Reference::referent_offset()) {
375 // Constant offset but not the reference_offset so just return
376 return;
377 }
378
379 // We only need to generate the runtime guards for instances.
380 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
381 if (btype != nullptr) {
382 if (btype->isa_aryptr()) {
383 // Array type so nothing to do
384 return;
385 }
386
387 const TypeInstPtr* itype = btype->isa_instptr();
388 if (itype != nullptr) {
389 // Can the klass of base_oop be statically determined to be
390 // _not_ a sub-class of Reference and _not_ Object?
391 ciKlass* klass = itype->instance_klass();
392 if (klass->is_loaded() &&
393 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
394 !kit->env()->Object_klass()->is_subtype_of(klass)) {
395 return;
396 }
397 }
398 }
399
400 // The compile time filters did not reject base_oop/offset so
401 // we need to generate the following runtime filters
402 //
403 // if (offset == java_lang_ref_Reference::_reference_offset) {
404 // if (instance_of(base, java.lang.ref.Reference)) {
405 // pre_barrier(_, pre_val, ...);
406 // }
407 // }
408
409 float likely = PROB_LIKELY( 0.999);
410 float unlikely = PROB_UNLIKELY(0.999);
411
412 IdealKit ideal(kit);
413
414 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset());
415
416 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
417 // Update graphKit memory and control from IdealKit.
418 kit->sync_kit(ideal);
419
420 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
421 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
422
423 // Update IdealKit memory and control from graphKit.
424 __ sync_kit(kit);
425
426 Node* one = __ ConI(1);
427 // is_instof == 0 if base_oop == nullptr
428 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
429
430 // Update graphKit from IdeakKit.
431 kit->sync_kit(ideal);
432
433 // Use the pre-barrier to record the value in the referent field
434 satb_write_barrier_pre(kit, false /* do_load */,
435 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */,
436 pre_val /* pre_val */,
437 T_OBJECT);
438 if (need_mem_bar) {
439 // Add memory barrier to prevent commoning reads from this field
440 // across safepoint since GC can change its value.
441 kit->insert_mem_bar(Op_MemBarCPUOrder);
442 }
443 // Update IdealKit from graphKit.
444 __ sync_kit(kit);
445
446 } __ end_if(); // _ref_type != ref_none
447 } __ end_if(); // offset == referent_offset
448
449 // Final sync IdealKit and GraphKit.
450 kit->final_sync(ideal);
451 }
452
453 #undef __
454
455 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() {
456 const Type **fields = TypeTuple::fields(2);
457 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
458 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
459 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
460
461 // create result type (range)
462 fields = TypeTuple::fields(0);
463 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
464
465 return TypeFunc::make(domain, range);
466 }
467
468 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() {
469 const Type **fields = TypeTuple::fields(1);
470 fields[TypeFunc::Parms+0] = TypeOopPtr::NOTNULL; // src oop
471 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
472
473 // create result type (range)
474 fields = TypeTuple::fields(0);
475 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
476
477 return TypeFunc::make(domain, range);
478 }
479
480 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type() {
481 const Type **fields = TypeTuple::fields(2);
482 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM; // original field value
483 fields[TypeFunc::Parms+1] = TypeRawPtr::BOTTOM; // original load address
484
485 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
486
487 // create result type (range)
488 fields = TypeTuple::fields(1);
489 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM;
490 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
491
492 return TypeFunc::make(domain, range);
493 }
494
495 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
496 DecoratorSet decorators = access.decorators();
497
498 const TypePtr* adr_type = access.addr().type();
499 Node* adr = access.addr().node();
500
501 if (!access.is_oop()) {
502 return BarrierSetC2::store_at_resolved(access, val);
503 }
504
505 if (access.is_parse_access()) {
506 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
507 GraphKit* kit = parse_access.kit();
508
509 uint adr_idx = kit->C->get_alias_index(adr_type);
510 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
511 Node* value = val.node();
512 value = shenandoah_iu_barrier(kit, value);
513 val.set_node(value);
514 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
515 static_cast<const TypeOopPtr*>(val.type()), nullptr /* pre_val */, access.type());
516 } else {
517 assert(access.is_opt_access(), "only for optimization passes");
518 assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code");
519 C2OptAccess& opt_access = static_cast<C2OptAccess&>(access);
520 PhaseGVN& gvn = opt_access.gvn();
521
522 if (ShenandoahIUBarrier) {
523 Node* enqueue = gvn.transform(new ShenandoahIUBarrierNode(val.node()));
524 val.set_node(enqueue);
525 }
526 }
527 return BarrierSetC2::store_at_resolved(access, val);
528 }
529
530 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
531 // 1: non-reference load, no additional barrier is needed
532 if (!access.is_oop()) {
533 return BarrierSetC2::load_at_resolved(access, val_type);;
534 }
535
536 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
537 DecoratorSet decorators = access.decorators();
538 BasicType type = access.type();
539
540 // 2: apply LRB if needed
541 if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
542 load = new ShenandoahLoadReferenceBarrierNode(nullptr, load, decorators);
543 if (access.is_parse_access()) {
544 load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load);
545 } else {
546 load = static_cast<C2OptAccess &>(access).gvn().transform(load);
547 }
548 }
549
550 // 3: apply keep-alive barrier for java.lang.ref.Reference if needed
551 if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
552 Node* top = Compile::current()->top();
553 Node* adr = access.addr().node();
554 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
555 Node* obj = access.base();
556
557 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
558 bool on_weak_ref = (decorators & (ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF)) != 0;
559 bool keep_alive = (decorators & AS_NO_KEEPALIVE) == 0;
560
561 // If we are reading the value of the referent field of a Reference
562 // object (either by using Unsafe directly or through reflection)
563 // then, if SATB is enabled, we need to record the referent in an
564 // SATB log buffer using the pre-barrier mechanism.
565 // Also we need to add memory barrier to prevent commoning reads
566 // from this field across safepoint since GC can change its value.
567 if (!on_weak_ref || (unknown && (offset == top || obj == top)) || !keep_alive) {
568 return load;
569 }
570
571 assert(access.is_parse_access(), "entry not supported at optimization time");
572 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
573 GraphKit* kit = parse_access.kit();
574 bool mismatched = (decorators & C2_MISMATCHED) != 0;
575 bool is_unordered = (decorators & MO_UNORDERED) != 0;
576 bool in_native = (decorators & IN_NATIVE) != 0;
577 bool need_cpu_mem_bar = !is_unordered || mismatched || in_native;
578
579 if (on_weak_ref) {
580 // Use the pre-barrier to record the value in the referent field
581 satb_write_barrier_pre(kit, false /* do_load */,
582 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */,
583 load /* pre_val */, T_OBJECT);
584 // Add memory barrier to prevent commoning reads from this field
585 // across safepoint since GC can change its value.
586 kit->insert_mem_bar(Op_MemBarCPUOrder);
587 } else if (unknown) {
588 // We do not require a mem bar inside pre_barrier if need_mem_bar
589 // is set: the barriers would be emitted by us.
590 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
591 }
592 }
593
594 return load;
595 }
596
597 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
598 Node* new_val, const Type* value_type) const {
599 GraphKit* kit = access.kit();
600 if (access.is_oop()) {
601 new_val = shenandoah_iu_barrier(kit, new_val);
602 shenandoah_write_barrier_pre(kit, false /* do_load */,
603 nullptr, nullptr, max_juint, nullptr, nullptr,
604 expected_val /* pre_val */, T_OBJECT);
605
606 MemNode::MemOrd mo = access.mem_node_mo();
607 Node* mem = access.memory();
608 Node* adr = access.addr().node();
609 const TypePtr* adr_type = access.addr().type();
610 Node* load_store = nullptr;
611
612 #ifdef _LP64
613 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
614 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
615 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
616 if (ShenandoahCASBarrier) {
617 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
618 } else {
619 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
620 }
621 } else
622 #endif
623 {
624 if (ShenandoahCASBarrier) {
625 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
626 } else {
627 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
628 }
629 }
630
631 access.set_raw_access(load_store);
632 pin_atomic_op(access);
633
634 #ifdef _LP64
635 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
636 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
637 }
638 #endif
639 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, load_store, access.decorators()));
640 return load_store;
641 }
642 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
643 }
644
645 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
646 Node* new_val, const Type* value_type) const {
647 GraphKit* kit = access.kit();
648 if (access.is_oop()) {
649 new_val = shenandoah_iu_barrier(kit, new_val);
650 shenandoah_write_barrier_pre(kit, false /* do_load */,
651 nullptr, nullptr, max_juint, nullptr, nullptr,
652 expected_val /* pre_val */, T_OBJECT);
653 DecoratorSet decorators = access.decorators();
654 MemNode::MemOrd mo = access.mem_node_mo();
655 Node* mem = access.memory();
656 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
657 Node* load_store = nullptr;
658 Node* adr = access.addr().node();
659 #ifdef _LP64
660 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
661 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
662 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
663 if (ShenandoahCASBarrier) {
664 if (is_weak_cas) {
665 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
666 } else {
667 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
668 }
669 } else {
670 if (is_weak_cas) {
671 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
672 } else {
673 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
674 }
675 }
676 } else
677 #endif
678 {
679 if (ShenandoahCASBarrier) {
680 if (is_weak_cas) {
681 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
682 } else {
683 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
684 }
685 } else {
686 if (is_weak_cas) {
687 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
688 } else {
689 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
690 }
691 }
692 }
693 access.set_raw_access(load_store);
694 pin_atomic_op(access);
695 return load_store;
696 }
697 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
698 }
699
700 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
701 GraphKit* kit = access.kit();
702 if (access.is_oop()) {
703 val = shenandoah_iu_barrier(kit, val);
704 }
705 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
706 if (access.is_oop()) {
707 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, result, access.decorators()));
708 shenandoah_write_barrier_pre(kit, false /* do_load */,
709 nullptr, nullptr, max_juint, nullptr, nullptr,
710 result /* pre_val */, T_OBJECT);
711 }
712 return result;
713 }
714
715
716 bool ShenandoahBarrierSetC2::is_gc_pre_barrier_node(Node* node) const {
717 return is_shenandoah_wb_pre_call(node);
718 }
719
720 // Support for GC barriers emitted during parsing
721 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
722 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier || node->Opcode() == Op_ShenandoahIUBarrier) return true;
723 if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) {
724 return false;
725 }
726 CallLeafNode *call = node->as_CallLeaf();
727 if (call->_name == nullptr) {
728 return false;
729 }
730
731 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 ||
732 strcmp(call->_name, "shenandoah_cas_obj") == 0 ||
733 strcmp(call->_name, "shenandoah_wb_pre") == 0;
734 }
735
736 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
737 if (c == nullptr) {
738 return c;
739 }
740 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
741 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
742 }
743 if (c->Opcode() == Op_ShenandoahIUBarrier) {
744 c = c->in(1);
745 }
746 return c;
747 }
748
749 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
750 return !ShenandoahBarrierC2Support::expand(C, igvn);
751 }
752
753 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
754 if (mode == LoopOptsShenandoahExpand) {
755 assert(UseShenandoahGC, "only for shenandoah");
756 ShenandoahBarrierC2Support::pin_and_expand(phase);
757 return true;
758 } else if (mode == LoopOptsShenandoahPostExpand) {
759 assert(UseShenandoahGC, "only for shenandoah");
760 visited.clear();
761 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
762 return true;
763 }
764 return false;
765 }
766
767 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, bool is_clone_instance, ArrayCopyPhase phase) const {
768 bool is_oop = is_reference_type(type);
769 if (!is_oop) {
770 return false;
771 }
772 if (ShenandoahSATBBarrier && tightly_coupled_alloc) {
773 if (phase == Optimization) {
774 return false;
775 }
776 return !is_clone;
777 }
778 if (phase == Optimization) {
779 return !ShenandoahIUBarrier;
780 }
781 return true;
782 }
783
784 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
785 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
786 if (src_type->isa_instptr() != nullptr) {
787 ciInstanceKlass* ik = src_type->is_instptr()->instance_klass();
788 if ((src_type->klass_is_exact() || !ik->has_subklass()) && !ik->has_injected_fields()) {
789 if (ik->has_object_fields()) {
790 return true;
791 } else {
792 if (!src_type->klass_is_exact()) {
793 Compile::current()->dependencies()->assert_leaf_type(ik);
794 }
795 }
796 } else {
797 return true;
798 }
799 } else if (src_type->isa_aryptr()) {
800 BasicType src_elem = src_type->isa_aryptr()->elem()->array_element_basic_type();
801 if (is_reference_type(src_elem, true)) {
802 return true;
803 }
804 } else {
805 return true;
806 }
807 return false;
808 }
809
810 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
811 Node* ctrl = ac->in(TypeFunc::Control);
812 Node* mem = ac->in(TypeFunc::Memory);
813 Node* src_base = ac->in(ArrayCopyNode::Src);
814 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
815 Node* dest_base = ac->in(ArrayCopyNode::Dest);
816 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
817 Node* length = ac->in(ArrayCopyNode::Length);
818
819 Node* src = phase->basic_plus_adr(src_base, src_offset);
820 Node* dest = phase->basic_plus_adr(dest_base, dest_offset);
821
822 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
823 // Check if heap is has forwarded objects. If it does, we need to call into the special
824 // routine that would fix up source references before we can continue.
825
826 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
827 Node* region = new RegionNode(PATH_LIMIT);
828 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM);
829
830 Node* thread = phase->transform_later(new ThreadLocalNode());
831 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
832 Node* gc_state_addr = phase->transform_later(new AddPNode(phase->C->top(), thread, offset));
833
834 uint gc_state_idx = Compile::AliasIdxRaw;
835 const TypePtr* gc_state_adr_type = nullptr; // debug-mode-only argument
836 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
837
838 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered));
839 int flags = ShenandoahHeap::HAS_FORWARDED;
840 if (ShenandoahIUBarrier) {
841 flags |= ShenandoahHeap::MARKING;
842 }
843 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(flags)));
844 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT)));
845 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne));
846
847 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If();
848 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff));
849 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff));
850
851 // Heap is stable, no need to do anything additional
852 region->init_req(_heap_stable, stable_ctrl);
853 mem_phi->init_req(_heap_stable, mem);
854
855 // Heap is unstable, call into clone barrier stub
856 Node* call = phase->make_leaf_call(unstable_ctrl, mem,
857 ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(),
858 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier),
859 "shenandoah_clone",
860 TypeRawPtr::BOTTOM,
861 src_base);
862 call = phase->transform_later(call);
863
864 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control));
865 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory));
866 region->init_req(_heap_unstable, ctrl);
867 mem_phi->init_req(_heap_unstable, mem);
868
869 // Wire up the actual arraycopy stub now
870 ctrl = phase->transform_later(region);
871 mem = phase->transform_later(mem_phi);
872
873 const char* name = "arraycopy";
874 call = phase->make_leaf_call(ctrl, mem,
875 OptoRuntime::fast_arraycopy_Type(),
876 phase->basictype2arraycopy(T_LONG, nullptr, nullptr, true, name, true),
877 name, TypeRawPtr::BOTTOM,
878 src, dest, length
879 LP64_ONLY(COMMA phase->top()));
880 call = phase->transform_later(call);
881
882 // Hook up the whole thing into the graph
883 phase->igvn().replace_node(ac, call);
884 } else {
885 BarrierSetC2::clone_at_expansion(phase, ac);
886 }
887 }
888
889
890 // Support for macro expanded GC barriers
891 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
892 if (node->Opcode() == Op_ShenandoahIUBarrier) {
893 state()->add_iu_barrier((ShenandoahIUBarrierNode*) node);
894 }
895 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
896 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
897 }
898 }
899
900 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
901 if (node->Opcode() == Op_ShenandoahIUBarrier) {
902 state()->remove_iu_barrier((ShenandoahIUBarrierNode*) node);
903 }
904 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
905 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
906 }
907 }
908
909 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const {
910 if (is_shenandoah_wb_pre_call(n)) {
911 shenandoah_eliminate_wb_pre(n, ¯o->igvn());
912 }
913 }
914
915 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
916 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
917 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
918 c = c->unique_ctrl_out();
919 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
920 c = c->unique_ctrl_out();
921 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
922 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
923 assert(iff->is_If(), "expect test");
924 if (!is_shenandoah_marking_if(igvn, iff)) {
925 c = c->unique_ctrl_out();
926 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
927 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
928 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
929 }
930 Node* cmpx = iff->in(1)->in(1);
931 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
932 igvn->rehash_node_delayed(call);
933 call->del_req(call->req()-1);
934 }
935
936 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
937 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
938 igvn->add_users_to_worklist(node);
939 }
940 }
941
942 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
943 for (uint i = 0; i < useful.size(); i++) {
944 Node* n = useful.at(i);
945 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
946 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
947 C->record_for_igvn(n->fast_out(i));
948 }
949 }
950 }
951 for (int i = state()->iu_barriers_count() - 1; i >= 0; i--) {
952 ShenandoahIUBarrierNode* n = state()->iu_barrier(i);
953 if (!useful.member(n)) {
954 state()->remove_iu_barrier(n);
955 }
956 }
957 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
958 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
959 if (!useful.member(n)) {
960 state()->remove_load_reference_barrier(n);
961 }
962 }
963 }
964
965 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
966 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
967 }
968
969 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
970 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
971 }
972
973 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
974 // expanded later, then now is the time to do so.
975 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
976
977 #ifdef ASSERT
978 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
979 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) {
980 ShenandoahBarrierC2Support::verify(Compile::current()->root());
981 } else if (phase == BarrierSetC2::BeforeCodeGen) {
982 // Verify Shenandoah pre-barriers
983 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset());
984
985 Unique_Node_List visited;
986 Node_List worklist;
987 // We're going to walk control flow backwards starting from the Root
988 worklist.push(compile->root());
989 while (worklist.size() > 0) {
990 Node *x = worklist.pop();
991 if (x == nullptr || x == compile->top()) continue;
992 if (visited.member(x)) {
993 continue;
994 } else {
995 visited.push(x);
996 }
997
998 if (x->is_Region()) {
999 for (uint i = 1; i < x->req(); i++) {
1000 worklist.push(x->in(i));
1001 }
1002 } else {
1003 worklist.push(x->in(0));
1004 // We are looking for the pattern:
1005 // /->ThreadLocal
1006 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
1007 // \->ConI(0)
1008 // We want to verify that the If and the LoadB have the same control
1009 // See GraphKit::g1_write_barrier_pre()
1010 if (x->is_If()) {
1011 IfNode *iff = x->as_If();
1012 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
1013 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
1014 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
1015 && cmp->in(1)->is_Load()) {
1016 LoadNode *load = cmp->in(1)->as_Load();
1017 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
1018 && load->in(2)->in(3)->is_Con()
1019 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
1020
1021 Node *if_ctrl = iff->in(0);
1022 Node *load_ctrl = load->in(0);
1023
1024 if (if_ctrl != load_ctrl) {
1025 // Skip possible CProj->NeverBranch in infinite loops
1026 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
1027 && if_ctrl->in(0)->is_NeverBranch()) {
1028 if_ctrl = if_ctrl->in(0)->in(0);
1029 }
1030 }
1031 assert(load_ctrl != nullptr && if_ctrl == load_ctrl, "controls must match");
1032 }
1033 }
1034 }
1035 }
1036 }
1037 }
1038 }
1039 }
1040 #endif
1041
1042 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
1043 if (is_shenandoah_wb_pre_call(n)) {
1044 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1045 if (n->req() > cnt) {
1046 Node* addp = n->in(cnt);
1047 if (has_only_shenandoah_wb_pre_uses(addp)) {
1048 n->del_req(cnt);
1049 if (can_reshape) {
1050 phase->is_IterGVN()->_worklist.push(addp);
1051 }
1052 return n;
1053 }
1054 }
1055 }
1056 if (n->Opcode() == Op_CmpP) {
1057 Node* in1 = n->in(1);
1058 Node* in2 = n->in(2);
1059
1060 // If one input is null, then step over the strong LRB barriers on the other input
1061 if (in1->bottom_type() == TypePtr::NULL_PTR &&
1062 !((in2->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1063 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in2)->decorators()))) {
1064 in2 = step_over_gc_barrier(in2);
1065 }
1066 if (in2->bottom_type() == TypePtr::NULL_PTR &&
1067 !((in1->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1068 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in1)->decorators()))) {
1069 in1 = step_over_gc_barrier(in1);
1070 }
1071
1072 if (in1 != n->in(1)) {
1073 n->set_req_X(1, in1, phase);
1074 assert(in2 == n->in(2), "only one change");
1075 return n;
1076 }
1077 if (in2 != n->in(2)) {
1078 n->set_req_X(2, in2, phase);
1079 return n;
1080 }
1081 } else if (can_reshape &&
1082 n->Opcode() == Op_If &&
1083 ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1084 n->in(0) != nullptr &&
1085 n->outcnt() == 2) {
1086 Node* dom = n->in(0);
1087 Node* prev_dom = n;
1088 int op = n->Opcode();
1089 int dist = 16;
1090 // Search up the dominator tree for another heap stable test
1091 while (dom->Opcode() != op || // Not same opcode?
1092 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1?
1093 prev_dom->in(0) != dom) { // One path of test does not dominate?
1094 if (dist < 0) return nullptr;
1095
1096 dist--;
1097 prev_dom = dom;
1098 dom = IfNode::up_one_dom(dom);
1099 if (!dom) return nullptr;
1100 }
1101
1102 // Check that we did not follow a loop back to ourselves
1103 if (n == dom) {
1104 return nullptr;
1105 }
1106
1107 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN());
1108 }
1109
1110 return nullptr;
1111 }
1112
1113 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1114 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1115 Node* u = n->fast_out(i);
1116 if (!is_shenandoah_wb_pre_call(u)) {
1117 return false;
1118 }
1119 }
1120 return n->outcnt() > 0;
1121 }
1122
1123 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode, Unique_Node_List& dead_nodes) const {
1124 switch (opcode) {
1125 case Op_CallLeaf:
1126 case Op_CallLeafNoFP: {
1127 assert (n->is_Call(), "");
1128 CallNode *call = n->as_Call();
1129 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1130 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1131 if (call->req() > cnt) {
1132 assert(call->req() == cnt + 1, "only one extra input");
1133 Node *addp = call->in(cnt);
1134 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1135 call->del_req(cnt);
1136 }
1137 }
1138 return false;
1139 }
1140 case Op_ShenandoahCompareAndSwapP:
1141 case Op_ShenandoahCompareAndSwapN:
1142 case Op_ShenandoahWeakCompareAndSwapN:
1143 case Op_ShenandoahWeakCompareAndSwapP:
1144 case Op_ShenandoahCompareAndExchangeP:
1145 case Op_ShenandoahCompareAndExchangeN:
1146 return true;
1147 case Op_ShenandoahLoadReferenceBarrier:
1148 assert(false, "should have been expanded already");
1149 return true;
1150 default:
1151 return false;
1152 }
1153 }
1154
1155 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1156 switch (opcode) {
1157 case Op_ShenandoahCompareAndExchangeP:
1158 case Op_ShenandoahCompareAndExchangeN:
1159 conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1160 // fallthrough
1161 case Op_ShenandoahWeakCompareAndSwapP:
1162 case Op_ShenandoahWeakCompareAndSwapN:
1163 case Op_ShenandoahCompareAndSwapP:
1164 case Op_ShenandoahCompareAndSwapN:
1165 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1166 return true;
1167 case Op_StoreP: {
1168 Node* adr = n->in(MemNode::Address);
1169 const Type* adr_type = gvn->type(adr);
1170 // Pointer stores in Shenandoah barriers looks like unsafe access.
1171 // Ignore such stores to be able scalar replace non-escaping
1172 // allocations.
1173 if (adr_type->isa_rawptr() && adr->is_AddP()) {
1174 Node* base = conn_graph->get_addp_base(adr);
1175 if (base->Opcode() == Op_LoadP &&
1176 base->in(MemNode::Address)->is_AddP()) {
1177 adr = base->in(MemNode::Address);
1178 Node* tls = conn_graph->get_addp_base(adr);
1179 if (tls->Opcode() == Op_ThreadLocal) {
1180 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1181 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1182 if (offs == buf_offset) {
1183 return true; // Pre barrier previous oop value store.
1184 }
1185 }
1186 }
1187 }
1188 return false;
1189 }
1190 case Op_ShenandoahIUBarrier:
1191 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1192 break;
1193 case Op_ShenandoahLoadReferenceBarrier:
1194 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist);
1195 return true;
1196 default:
1197 // Nothing
1198 break;
1199 }
1200 return false;
1201 }
1202
1203 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1204 switch (opcode) {
1205 case Op_ShenandoahCompareAndExchangeP:
1206 case Op_ShenandoahCompareAndExchangeN: {
1207 Node *adr = n->in(MemNode::Address);
1208 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, nullptr);
1209 // fallthrough
1210 }
1211 case Op_ShenandoahCompareAndSwapP:
1212 case Op_ShenandoahCompareAndSwapN:
1213 case Op_ShenandoahWeakCompareAndSwapP:
1214 case Op_ShenandoahWeakCompareAndSwapN:
1215 return conn_graph->add_final_edges_unsafe_access(n, opcode);
1216 case Op_ShenandoahIUBarrier:
1217 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), nullptr);
1218 return true;
1219 case Op_ShenandoahLoadReferenceBarrier:
1220 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), nullptr);
1221 return true;
1222 default:
1223 // Nothing
1224 break;
1225 }
1226 return false;
1227 }
1228
1229 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1230 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1231 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1232
1233 }
1234
1235 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1236 switch (opcode) {
1237 case Op_ShenandoahCompareAndExchangeP:
1238 case Op_ShenandoahCompareAndExchangeN:
1239 case Op_ShenandoahWeakCompareAndSwapP:
1240 case Op_ShenandoahWeakCompareAndSwapN:
1241 case Op_ShenandoahCompareAndSwapP:
1242 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree
1243 Node* newval = n->in(MemNode::ValueIn);
1244 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1245 Node* pair = new BinaryNode(oldval, newval);
1246 n->set_req(MemNode::ValueIn,pair);
1247 n->del_req(LoadStoreConditionalNode::ExpectedIn);
1248 return true;
1249 }
1250 default:
1251 break;
1252 }
1253 return false;
1254 }
1255
1256 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1257 return xop == Op_ShenandoahCompareAndExchangeP ||
1258 xop == Op_ShenandoahCompareAndExchangeN ||
1259 xop == Op_ShenandoahWeakCompareAndSwapP ||
1260 xop == Op_ShenandoahWeakCompareAndSwapN ||
1261 xop == Op_ShenandoahCompareAndSwapN ||
1262 xop == Op_ShenandoahCompareAndSwapP;
1263 }