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