1 /*
2 * Copyright (c) 2018, 2021, 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, NULL)),
52 _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8, 0, NULL)) {
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, PhaseTransform* 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, phase);
105
106 if (offset == Type::OffsetBot) {
107 return false; // cannot unalias unless there are precise offsets
108 }
109
110 if (alloc == NULL) {
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 == NULL) {
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, phase),
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 == NULL || 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(obj != NULL, "must have a base");
203 assert(adr != NULL, "where are loading from?");
204 assert(pre_val == NULL, "loaded already?");
205 assert(val_type != NULL, "need a type");
206
207 if (ReduceInitialCardMarks
208 && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
209 return;
210 }
211
212 } else {
213 // In this case both val_type and alias_idx are unused.
214 assert(pre_val != NULL, "must be loaded already");
215 // Nothing to be done if pre_val is null.
216 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
217 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
218 }
219 assert(bt == T_OBJECT, "or we shouldn't be here");
220
221 IdealKit ideal(kit, true);
222
223 Node* tls = __ thread(); // ThreadLocalStorage
224
225 Node* no_base = __ top();
226 Node* zero = __ ConI(0);
227 Node* zeroX = __ ConX(0);
228
229 float likely = PROB_LIKELY(0.999);
230 float unlikely = PROB_UNLIKELY(0.999);
231
232 // Offsets into the thread
233 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
234 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
235
236 // Now the actual pointers into the thread
237 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
238 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
239
240 // Now some of the values
241 Node* marking;
242 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())));
243 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
244 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
245 assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape");
246
247 // if (!marking)
248 __ if_then(marking, BoolTest::ne, zero, unlikely); {
249 BasicType index_bt = TypeX_X->basic_type();
250 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading Shenandoah SATBMarkQueue::_index with wrong size.");
251 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
252
253 if (do_load) {
254 // load original value
255 // alias_idx correct??
256 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
257 }
258
259 // if (pre_val != NULL)
260 __ if_then(pre_val, BoolTest::ne, kit->null()); {
261 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
262
263 // is the queue for this thread full?
264 __ if_then(index, BoolTest::ne, zeroX, likely); {
265
266 // decrement the index
267 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
268
269 // Now get the buffer location we will log the previous value into and store it
270 Node *log_addr = __ AddP(no_base, buffer, next_index);
271 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
272 // update the index
273 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
274
275 } __ else_(); {
276
277 // logging buffer is full, call the runtime
278 const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type();
279 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls);
280 } __ end_if(); // (!index)
281 } __ end_if(); // (pre_val != NULL)
282 } __ end_if(); // (!marking)
283
284 // Final sync IdealKit and GraphKit.
285 kit->final_sync(ideal);
286
287 if (ShenandoahSATBBarrier && adr != NULL) {
288 Node* c = kit->control();
289 Node* call = c->in(1)->in(1)->in(1)->in(0);
290 assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected");
291 call->add_req(adr);
292 }
293 }
294
295 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) {
296 return call->is_CallLeaf() &&
297 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry);
298 }
299
300 bool ShenandoahBarrierSetC2::is_shenandoah_lrb_call(Node* call) {
301 if (!call->is_CallLeaf()) {
302 return false;
303 }
304
305 address entry_point = call->as_CallLeaf()->entry_point();
306 return (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong)) ||
307 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow)) ||
308 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak)) ||
309 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow)) ||
310 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom));
311 }
312
313 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseTransform *phase, Node* n) {
314 if (n->Opcode() != Op_If) {
315 return false;
316 }
317
318 Node* bol = n->in(1);
319 assert(bol->is_Bool(), "");
320 Node* cmpx = bol->in(1);
321 if (bol->as_Bool()->_test._test == BoolTest::ne &&
322 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
323 is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
324 cmpx->in(1)->in(2)->is_Con() &&
325 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
326 return true;
327 }
328
329 return false;
330 }
331
332 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
333 if (!n->is_Load()) return false;
334 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
335 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
336 && n->in(2)->in(3)->is_Con()
337 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
338 }
339
340 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
341 bool do_load,
342 Node* obj,
343 Node* adr,
344 uint alias_idx,
345 Node* val,
346 const TypeOopPtr* val_type,
347 Node* pre_val,
348 BasicType bt) const {
349 if (ShenandoahSATBBarrier) {
350 IdealKit ideal(kit);
351 kit->sync_kit(ideal);
352
353 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
354
355 ideal.sync_kit(kit);
356 kit->final_sync(ideal);
357 }
358 }
359
360 // Helper that guards and inserts a pre-barrier.
361 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
362 Node* pre_val, bool need_mem_bar) const {
363 // We could be accessing the referent field of a reference object. If so, when Shenandoah
364 // is enabled, we need to log the value in the referent field in an SATB buffer.
365 // This routine performs some compile time filters and generates suitable
366 // runtime filters that guard the pre-barrier code.
367 // Also add memory barrier for non volatile load from the referent field
368 // to prevent commoning of loads across safepoint.
369
370 // Some compile time checks.
371
372 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
373 const TypeX* otype = offset->find_intptr_t_type();
374 if (otype != NULL && otype->is_con() &&
375 otype->get_con() != java_lang_ref_Reference::referent_offset()) {
376 // Constant offset but not the reference_offset so just return
377 return;
378 }
379
380 // We only need to generate the runtime guards for instances.
381 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
382 if (btype != NULL) {
383 if (btype->isa_aryptr()) {
384 // Array type so nothing to do
385 return;
386 }
387
388 const TypeInstPtr* itype = btype->isa_instptr();
389 if (itype != NULL) {
390 // Can the klass of base_oop be statically determined to be
391 // _not_ a sub-class of Reference and _not_ Object?
392 ciKlass* klass = itype->klass();
393 if ( klass->is_loaded() &&
394 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
395 !kit->env()->Object_klass()->is_subtype_of(klass)) {
396 return;
397 }
398 }
399 }
400
401 // The compile time filters did not reject base_oop/offset so
402 // we need to generate the following runtime filters
403 //
404 // if (offset == java_lang_ref_Reference::_reference_offset) {
405 // if (instance_of(base, java.lang.ref.Reference)) {
406 // pre_barrier(_, pre_val, ...);
407 // }
408 // }
409
410 float likely = PROB_LIKELY( 0.999);
411 float unlikely = PROB_UNLIKELY(0.999);
412
413 IdealKit ideal(kit);
414
415 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset());
416
417 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
418 // Update graphKit memory and control from IdealKit.
419 kit->sync_kit(ideal);
420
421 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
422 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
423
424 // Update IdealKit memory and control from graphKit.
425 __ sync_kit(kit);
426
427 Node* one = __ ConI(1);
428 // is_instof == 0 if base_oop == NULL
429 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
430
431 // Update graphKit from IdeakKit.
432 kit->sync_kit(ideal);
433
434 // Use the pre-barrier to record the value in the referent field
435 satb_write_barrier_pre(kit, false /* do_load */,
436 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
437 pre_val /* pre_val */,
438 T_OBJECT);
439 if (need_mem_bar) {
440 // Add memory barrier to prevent commoning reads from this field
441 // across safepoint since GC can change its value.
442 kit->insert_mem_bar(Op_MemBarCPUOrder);
443 }
444 // Update IdealKit from graphKit.
445 __ sync_kit(kit);
446
447 } __ end_if(); // _ref_type != ref_none
448 } __ end_if(); // offset == referent_offset
449
450 // Final sync IdealKit and GraphKit.
451 kit->final_sync(ideal);
452 }
453
454 #undef __
455
456 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() {
457 const Type **fields = TypeTuple::fields(2);
458 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
459 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
460 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
461
462 // create result type (range)
463 fields = TypeTuple::fields(0);
464 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
465
466 return TypeFunc::make(domain, range);
467 }
468
469 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() {
470 const Type **fields = TypeTuple::fields(1);
471 fields[TypeFunc::Parms+0] = TypeOopPtr::NOTNULL; // src oop
472 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
473
474 // create result type (range)
475 fields = TypeTuple::fields(0);
476 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
477
478 return TypeFunc::make(domain, range);
479 }
480
481 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type() {
482 const Type **fields = TypeTuple::fields(2);
483 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM; // original field value
484 fields[TypeFunc::Parms+1] = TypeRawPtr::BOTTOM; // original load address
485
486 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
487
488 // create result type (range)
489 fields = TypeTuple::fields(1);
490 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM;
491 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
492
493 return TypeFunc::make(domain, range);
494 }
495
496 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
497 DecoratorSet decorators = access.decorators();
498
499 const TypePtr* adr_type = access.addr().type();
500 Node* adr = access.addr().node();
501
502 bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
503 bool on_heap = (decorators & IN_HEAP) != 0;
504
505 if (!access.is_oop() || (!on_heap && !anonymous)) {
506 return BarrierSetC2::store_at_resolved(access, val);
507 }
508
509 if (access.is_parse_access()) {
510 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
511 GraphKit* kit = parse_access.kit();
512
513 uint adr_idx = kit->C->get_alias_index(adr_type);
514 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
515 Node* value = val.node();
516 value = shenandoah_iu_barrier(kit, value);
517 val.set_node(value);
518 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
519 static_cast<const TypeOopPtr*>(val.type()), NULL /* pre_val */, access.type());
520 } else {
521 assert(access.is_opt_access(), "only for optimization passes");
522 assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code");
523 C2OptAccess& opt_access = static_cast<C2OptAccess&>(access);
524 PhaseGVN& gvn = opt_access.gvn();
525
526 if (ShenandoahIUBarrier) {
527 Node* enqueue = gvn.transform(new ShenandoahIUBarrierNode(val.node()));
528 val.set_node(enqueue);
529 }
530 }
531 return BarrierSetC2::store_at_resolved(access, val);
532 }
533
534 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
535 // 1: non-reference load, no additional barrier is needed
536 if (!access.is_oop()) {
537 return BarrierSetC2::load_at_resolved(access, val_type);;
538 }
539
540 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
541 DecoratorSet decorators = access.decorators();
542 BasicType type = access.type();
543
544 // 2: apply LRB if needed
545 if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
546 load = new ShenandoahLoadReferenceBarrierNode(NULL, load, decorators);
547 if (access.is_parse_access()) {
548 load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load);
549 } else {
550 load = static_cast<C2OptAccess &>(access).gvn().transform(load);
551 }
552 }
553
554 // 3: apply keep-alive barrier for java.lang.ref.Reference if needed
555 if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
556 Node* top = Compile::current()->top();
557 Node* adr = access.addr().node();
558 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
559 Node* obj = access.base();
560
561 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
562 bool on_weak_ref = (decorators & (ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF)) != 0;
563 bool keep_alive = (decorators & AS_NO_KEEPALIVE) == 0;
564
565 // If we are reading the value of the referent field of a Reference
566 // object (either by using Unsafe directly or through reflection)
567 // then, if SATB is enabled, we need to record the referent in an
568 // SATB log buffer using the pre-barrier mechanism.
569 // Also we need to add memory barrier to prevent commoning reads
570 // from this field across safepoint since GC can change its value.
571 if (!on_weak_ref || (unknown && (offset == top || obj == top)) || !keep_alive) {
572 return load;
573 }
574
575 assert(access.is_parse_access(), "entry not supported at optimization time");
576 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
577 GraphKit* kit = parse_access.kit();
578 bool mismatched = (decorators & C2_MISMATCHED) != 0;
579 bool is_unordered = (decorators & MO_UNORDERED) != 0;
580 bool in_native = (decorators & IN_NATIVE) != 0;
581 bool need_cpu_mem_bar = !is_unordered || mismatched || in_native;
582
583 if (on_weak_ref) {
584 // Use the pre-barrier to record the value in the referent field
585 satb_write_barrier_pre(kit, false /* do_load */,
586 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
587 load /* pre_val */, T_OBJECT);
588 // Add memory barrier to prevent commoning reads from this field
589 // across safepoint since GC can change its value.
590 kit->insert_mem_bar(Op_MemBarCPUOrder);
591 } else if (unknown) {
592 // We do not require a mem bar inside pre_barrier if need_mem_bar
593 // is set: the barriers would be emitted by us.
594 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
595 }
596 }
597
598 return load;
599 }
600
601 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
602 Node* new_val, const Type* value_type) const {
603 GraphKit* kit = access.kit();
604 if (access.is_oop()) {
605 new_val = shenandoah_iu_barrier(kit, new_val);
606 shenandoah_write_barrier_pre(kit, false /* do_load */,
607 NULL, NULL, max_juint, NULL, NULL,
608 expected_val /* pre_val */, T_OBJECT);
609
610 MemNode::MemOrd mo = access.mem_node_mo();
611 Node* mem = access.memory();
612 Node* adr = access.addr().node();
613 const TypePtr* adr_type = access.addr().type();
614 Node* load_store = NULL;
615
616 #ifdef _LP64
617 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
618 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
619 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
620 if (ShenandoahCASBarrier) {
621 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
622 } else {
623 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
624 }
625 } else
626 #endif
627 {
628 if (ShenandoahCASBarrier) {
629 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
630 } else {
631 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
632 }
633 }
634
635 access.set_raw_access(load_store);
636 pin_atomic_op(access);
637
638 #ifdef _LP64
639 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
640 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
641 }
642 #endif
643 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, load_store, access.decorators()));
644 return load_store;
645 }
646 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
647 }
648
649 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
650 Node* new_val, const Type* value_type) const {
651 GraphKit* kit = access.kit();
652 if (access.is_oop()) {
653 new_val = shenandoah_iu_barrier(kit, new_val);
654 shenandoah_write_barrier_pre(kit, false /* do_load */,
655 NULL, NULL, max_juint, NULL, NULL,
656 expected_val /* pre_val */, T_OBJECT);
657 DecoratorSet decorators = access.decorators();
658 MemNode::MemOrd mo = access.mem_node_mo();
659 Node* mem = access.memory();
660 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
661 Node* load_store = NULL;
662 Node* adr = access.addr().node();
663 #ifdef _LP64
664 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
665 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
666 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
667 if (ShenandoahCASBarrier) {
668 if (is_weak_cas) {
669 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
670 } else {
671 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
672 }
673 } else {
674 if (is_weak_cas) {
675 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
676 } else {
677 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
678 }
679 }
680 } else
681 #endif
682 {
683 if (ShenandoahCASBarrier) {
684 if (is_weak_cas) {
685 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
686 } else {
687 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
688 }
689 } else {
690 if (is_weak_cas) {
691 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
692 } else {
693 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
694 }
695 }
696 }
697 access.set_raw_access(load_store);
698 pin_atomic_op(access);
699 return load_store;
700 }
701 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
702 }
703
704 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
705 GraphKit* kit = access.kit();
706 if (access.is_oop()) {
707 val = shenandoah_iu_barrier(kit, val);
708 }
709 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
710 if (access.is_oop()) {
711 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, result, access.decorators()));
712 shenandoah_write_barrier_pre(kit, false /* do_load */,
713 NULL, NULL, max_juint, NULL, NULL,
714 result /* pre_val */, T_OBJECT);
715 }
716 return result;
717 }
718
719 // Support for GC barriers emitted during parsing
720 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
721 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) return true;
722 if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) {
723 return false;
724 }
725 CallLeafNode *call = node->as_CallLeaf();
726 if (call->_name == NULL) {
727 return false;
728 }
729
730 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 ||
731 strcmp(call->_name, "shenandoah_cas_obj") == 0 ||
732 strcmp(call->_name, "shenandoah_wb_pre") == 0;
733 }
734
735 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
736 if (c == NULL) {
737 return c;
738 }
739 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
740 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
741 }
742 if (c->Opcode() == Op_ShenandoahIUBarrier) {
743 c = c->in(1);
744 }
745 return c;
746 }
747
748 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
749 return !ShenandoahBarrierC2Support::expand(C, igvn);
750 }
751
752 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
753 if (mode == LoopOptsShenandoahExpand) {
754 assert(UseShenandoahGC, "only for shenandoah");
755 ShenandoahBarrierC2Support::pin_and_expand(phase);
756 return true;
757 } else if (mode == LoopOptsShenandoahPostExpand) {
758 assert(UseShenandoahGC, "only for shenandoah");
759 visited.clear();
760 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
761 return true;
762 }
763 return false;
764 }
765
766 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, bool is_clone_instance, ArrayCopyPhase phase) const {
767 bool is_oop = is_reference_type(type);
768 if (!is_oop) {
769 return false;
770 }
771 if (ShenandoahSATBBarrier && tightly_coupled_alloc) {
772 if (phase == Optimization) {
773 return false;
774 }
775 return !is_clone;
776 }
777 if (phase == Optimization) {
778 return !ShenandoahIUBarrier;
779 }
780 return true;
781 }
782
783 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
784 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
785 if (src_type->isa_instptr() != NULL) {
786 ciInstanceKlass* ik = src_type->klass()->as_instance_klass();
787 if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) {
788 if (ik->has_object_fields()) {
789 return true;
790 } else {
791 if (!src_type->klass_is_exact()) {
792 Compile::current()->dependencies()->assert_leaf_type(ik);
793 }
794 }
795 } else {
796 return true;
797 }
798 } else if (src_type->isa_aryptr()) {
799 BasicType src_elem = src_type->klass()->as_array_klass()->element_type()->basic_type();
800 if (is_reference_type(src_elem)) {
801 return true;
802 }
803 } else {
804 return true;
805 }
806 return false;
807 }
808
809 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
810 Node* ctrl = ac->in(TypeFunc::Control);
811 Node* mem = ac->in(TypeFunc::Memory);
812 Node* src_base = ac->in(ArrayCopyNode::Src);
813 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
814 Node* dest_base = ac->in(ArrayCopyNode::Dest);
815 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
816 Node* length = ac->in(ArrayCopyNode::Length);
817
818 Node* src = phase->basic_plus_adr(src_base, src_offset);
819 Node* dest = phase->basic_plus_adr(dest_base, dest_offset);
820
821 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
822 // Check if heap is has forwarded objects. If it does, we need to call into the special
823 // routine that would fix up source references before we can continue.
824
825 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
826 Node* region = new RegionNode(PATH_LIMIT);
827 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM);
828
829 Node* thread = phase->transform_later(new ThreadLocalNode());
830 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
831 Node* gc_state_addr = phase->transform_later(new AddPNode(phase->C->top(), thread, offset));
832
833 uint gc_state_idx = Compile::AliasIdxRaw;
834 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument
835 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
836
837 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered));
838 int flags = ShenandoahHeap::HAS_FORWARDED;
839 if (ShenandoahIUBarrier) {
840 flags |= ShenandoahHeap::MARKING;
841 }
842 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(flags)));
843 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT)));
844 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne));
845
846 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If();
847 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff));
848 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff));
849
850 // Heap is stable, no need to do anything additional
851 region->init_req(_heap_stable, stable_ctrl);
852 mem_phi->init_req(_heap_stable, mem);
853
854 // Heap is unstable, call into clone barrier stub
855 Node* call = phase->make_leaf_call(unstable_ctrl, mem,
856 ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(),
857 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier),
858 "shenandoah_clone",
859 TypeRawPtr::BOTTOM,
860 src_base);
861 call = phase->transform_later(call);
862
863 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control));
864 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory));
865 region->init_req(_heap_unstable, ctrl);
866 mem_phi->init_req(_heap_unstable, mem);
867
868 // Wire up the actual arraycopy stub now
869 ctrl = phase->transform_later(region);
870 mem = phase->transform_later(mem_phi);
871
872 const char* name = "arraycopy";
873 call = phase->make_leaf_call(ctrl, mem,
874 OptoRuntime::fast_arraycopy_Type(),
875 phase->basictype2arraycopy(T_LONG, NULL, NULL, true, name, true),
876 name, TypeRawPtr::BOTTOM,
877 src, dest, length
878 LP64_ONLY(COMMA phase->top()));
879 call = phase->transform_later(call);
880
881 // Hook up the whole thing into the graph
882 phase->igvn().replace_node(ac, call);
883 } else {
884 BarrierSetC2::clone_at_expansion(phase, ac);
885 }
886 }
887
888
889 // Support for macro expanded GC barriers
890 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
891 if (node->Opcode() == Op_ShenandoahIUBarrier) {
892 state()->add_iu_barrier((ShenandoahIUBarrierNode*) node);
893 }
894 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
895 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
896 }
897 }
898
899 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
900 if (node->Opcode() == Op_ShenandoahIUBarrier) {
901 state()->remove_iu_barrier((ShenandoahIUBarrierNode*) node);
902 }
903 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
904 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
905 }
906 }
907
908 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const {
909 if (is_shenandoah_wb_pre_call(n)) {
910 shenandoah_eliminate_wb_pre(n, ¯o->igvn());
911 }
912 }
913
914 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
915 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
916 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
917 c = c->unique_ctrl_out();
918 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
919 c = c->unique_ctrl_out();
920 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
921 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
922 assert(iff->is_If(), "expect test");
923 if (!is_shenandoah_marking_if(igvn, iff)) {
924 c = c->unique_ctrl_out();
925 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
926 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
927 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
928 }
929 Node* cmpx = iff->in(1)->in(1);
930 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
931 igvn->rehash_node_delayed(call);
932 call->del_req(call->req()-1);
933 }
934
935 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
936 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
937 igvn->add_users_to_worklist(node);
938 }
939 }
940
941 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
942 for (uint i = 0; i < useful.size(); i++) {
943 Node* n = useful.at(i);
944 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
945 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
946 C->record_for_igvn(n->fast_out(i));
947 }
948 }
949 }
950 for (int i = state()->iu_barriers_count() - 1; i >= 0; i--) {
951 ShenandoahIUBarrierNode* n = state()->iu_barrier(i);
952 if (!useful.member(n)) {
953 state()->remove_iu_barrier(n);
954 }
955 }
956 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
957 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
958 if (!useful.member(n)) {
959 state()->remove_load_reference_barrier(n);
960 }
961 }
962 }
963
964 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
965 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
966 }
967
968 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
969 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
970 }
971
972 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
973 // expanded later, then now is the time to do so.
974 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
975
976 #ifdef ASSERT
977 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
978 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) {
979 ShenandoahBarrierC2Support::verify(Compile::current()->root());
980 } else if (phase == BarrierSetC2::BeforeCodeGen) {
981 // Verify Shenandoah pre-barriers
982 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset());
983
984 Unique_Node_List visited;
985 Node_List worklist;
986 // We're going to walk control flow backwards starting from the Root
987 worklist.push(compile->root());
988 while (worklist.size() > 0) {
989 Node *x = worklist.pop();
990 if (x == NULL || x == compile->top()) continue;
991 if (visited.member(x)) {
992 continue;
993 } else {
994 visited.push(x);
995 }
996
997 if (x->is_Region()) {
998 for (uint i = 1; i < x->req(); i++) {
999 worklist.push(x->in(i));
1000 }
1001 } else {
1002 worklist.push(x->in(0));
1003 // We are looking for the pattern:
1004 // /->ThreadLocal
1005 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
1006 // \->ConI(0)
1007 // We want to verify that the If and the LoadB have the same control
1008 // See GraphKit::g1_write_barrier_pre()
1009 if (x->is_If()) {
1010 IfNode *iff = x->as_If();
1011 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
1012 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
1013 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
1014 && cmp->in(1)->is_Load()) {
1015 LoadNode *load = cmp->in(1)->as_Load();
1016 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
1017 && load->in(2)->in(3)->is_Con()
1018 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
1019
1020 Node *if_ctrl = iff->in(0);
1021 Node *load_ctrl = load->in(0);
1022
1023 if (if_ctrl != load_ctrl) {
1024 // Skip possible CProj->NeverBranch in infinite loops
1025 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
1026 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
1027 if_ctrl = if_ctrl->in(0)->in(0);
1028 }
1029 }
1030 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
1031 }
1032 }
1033 }
1034 }
1035 }
1036 }
1037 }
1038 }
1039 #endif
1040
1041 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
1042 if (is_shenandoah_wb_pre_call(n)) {
1043 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1044 if (n->req() > cnt) {
1045 Node* addp = n->in(cnt);
1046 if (has_only_shenandoah_wb_pre_uses(addp)) {
1047 n->del_req(cnt);
1048 if (can_reshape) {
1049 phase->is_IterGVN()->_worklist.push(addp);
1050 }
1051 return n;
1052 }
1053 }
1054 }
1055 if (n->Opcode() == Op_CmpP) {
1056 Node* in1 = n->in(1);
1057 Node* in2 = n->in(2);
1058
1059 // If one input is NULL, then step over the strong LRB barriers on the other input
1060 if (in1->bottom_type() == TypePtr::NULL_PTR &&
1061 !((in2->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1062 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in2)->decorators()))) {
1063 in2 = step_over_gc_barrier(in2);
1064 }
1065 if (in2->bottom_type() == TypePtr::NULL_PTR &&
1066 !((in1->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1067 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in1)->decorators()))) {
1068 in1 = step_over_gc_barrier(in1);
1069 }
1070
1071 if (in1 != n->in(1)) {
1072 n->set_req_X(1, in1, phase);
1073 assert(in2 == n->in(2), "only one change");
1074 return n;
1075 }
1076 if (in2 != n->in(2)) {
1077 n->set_req_X(2, in2, phase);
1078 return n;
1079 }
1080 } else if (can_reshape &&
1081 n->Opcode() == Op_If &&
1082 ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1083 n->in(0) != NULL) {
1084 Node* dom = n->in(0);
1085 Node* prev_dom = n;
1086 int op = n->Opcode();
1087 int dist = 16;
1088 // Search up the dominator tree for another heap stable test
1089 while (dom->Opcode() != op || // Not same opcode?
1090 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1?
1091 prev_dom->in(0) != dom) { // One path of test does not dominate?
1092 if (dist < 0) return NULL;
1093
1094 dist--;
1095 prev_dom = dom;
1096 dom = IfNode::up_one_dom(dom);
1097 if (!dom) return NULL;
1098 }
1099
1100 // Check that we did not follow a loop back to ourselves
1101 if (n == dom) {
1102 return NULL;
1103 }
1104
1105 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN());
1106 }
1107
1108 return NULL;
1109 }
1110
1111 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1112 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1113 Node* u = n->fast_out(i);
1114 if (!is_shenandoah_wb_pre_call(u)) {
1115 return false;
1116 }
1117 }
1118 return n->outcnt() > 0;
1119 }
1120
1121 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode) const {
1122 switch (opcode) {
1123 case Op_CallLeaf:
1124 case Op_CallLeafNoFP: {
1125 assert (n->is_Call(), "");
1126 CallNode *call = n->as_Call();
1127 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1128 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1129 if (call->req() > cnt) {
1130 assert(call->req() == cnt + 1, "only one extra input");
1131 Node *addp = call->in(cnt);
1132 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1133 call->del_req(cnt);
1134 }
1135 }
1136 return false;
1137 }
1138 case Op_ShenandoahCompareAndSwapP:
1139 case Op_ShenandoahCompareAndSwapN:
1140 case Op_ShenandoahWeakCompareAndSwapN:
1141 case Op_ShenandoahWeakCompareAndSwapP:
1142 case Op_ShenandoahCompareAndExchangeP:
1143 case Op_ShenandoahCompareAndExchangeN:
1144 return true;
1145 case Op_ShenandoahLoadReferenceBarrier:
1146 assert(false, "should have been expanded already");
1147 return true;
1148 default:
1149 return false;
1150 }
1151 }
1152
1153 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1154 switch (opcode) {
1155 case Op_ShenandoahCompareAndExchangeP:
1156 case Op_ShenandoahCompareAndExchangeN:
1157 conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1158 // fallthrough
1159 case Op_ShenandoahWeakCompareAndSwapP:
1160 case Op_ShenandoahWeakCompareAndSwapN:
1161 case Op_ShenandoahCompareAndSwapP:
1162 case Op_ShenandoahCompareAndSwapN:
1163 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1164 return true;
1165 case Op_StoreP: {
1166 Node* adr = n->in(MemNode::Address);
1167 const Type* adr_type = gvn->type(adr);
1168 // Pointer stores in Shenandoah barriers looks like unsafe access.
1169 // Ignore such stores to be able scalar replace non-escaping
1170 // allocations.
1171 if (adr_type->isa_rawptr() && adr->is_AddP()) {
1172 Node* base = conn_graph->get_addp_base(adr);
1173 if (base->Opcode() == Op_LoadP &&
1174 base->in(MemNode::Address)->is_AddP()) {
1175 adr = base->in(MemNode::Address);
1176 Node* tls = conn_graph->get_addp_base(adr);
1177 if (tls->Opcode() == Op_ThreadLocal) {
1178 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1179 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1180 if (offs == buf_offset) {
1181 return true; // Pre barrier previous oop value store.
1182 }
1183 }
1184 }
1185 }
1186 return false;
1187 }
1188 case Op_ShenandoahIUBarrier:
1189 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1190 break;
1191 case Op_ShenandoahLoadReferenceBarrier:
1192 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist);
1193 return true;
1194 default:
1195 // Nothing
1196 break;
1197 }
1198 return false;
1199 }
1200
1201 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1202 switch (opcode) {
1203 case Op_ShenandoahCompareAndExchangeP:
1204 case Op_ShenandoahCompareAndExchangeN: {
1205 Node *adr = n->in(MemNode::Address);
1206 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
1207 // fallthrough
1208 }
1209 case Op_ShenandoahCompareAndSwapP:
1210 case Op_ShenandoahCompareAndSwapN:
1211 case Op_ShenandoahWeakCompareAndSwapP:
1212 case Op_ShenandoahWeakCompareAndSwapN:
1213 return conn_graph->add_final_edges_unsafe_access(n, opcode);
1214 case Op_ShenandoahIUBarrier:
1215 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), NULL);
1216 return true;
1217 case Op_ShenandoahLoadReferenceBarrier:
1218 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), NULL);
1219 return true;
1220 default:
1221 // Nothing
1222 break;
1223 }
1224 return false;
1225 }
1226
1227 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1228 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1229 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1230
1231 }
1232
1233 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1234 switch (opcode) {
1235 case Op_ShenandoahCompareAndExchangeP:
1236 case Op_ShenandoahCompareAndExchangeN:
1237 case Op_ShenandoahWeakCompareAndSwapP:
1238 case Op_ShenandoahWeakCompareAndSwapN:
1239 case Op_ShenandoahCompareAndSwapP:
1240 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree
1241 Node* newval = n->in(MemNode::ValueIn);
1242 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1243 Node* pair = new BinaryNode(oldval, newval);
1244 n->set_req(MemNode::ValueIn,pair);
1245 n->del_req(LoadStoreConditionalNode::ExpectedIn);
1246 return true;
1247 }
1248 default:
1249 break;
1250 }
1251 return false;
1252 }
1253
1254 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1255 return xop == Op_ShenandoahCompareAndExchangeP ||
1256 xop == Op_ShenandoahCompareAndExchangeN ||
1257 xop == Op_ShenandoahWeakCompareAndSwapP ||
1258 xop == Op_ShenandoahWeakCompareAndSwapN ||
1259 xop == Op_ShenandoahCompareAndSwapN ||
1260 xop == Op_ShenandoahCompareAndSwapP;
1261 }