1 /*
2 * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved.
3 *
4 * This code is free software; you can redistribute it and/or modify it
5 * under the terms of the GNU General Public License version 2 only, as
6 * published by the Free Software Foundation.
7 *
8 * This code is distributed in the hope that it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
11 * version 2 for more details (a copy is included in the LICENSE file that
12 * accompanied this code).
13 *
14 * You should have received a copy of the GNU General Public License version
15 * 2 along with this work; if not, write to the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
17 *
18 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
19 * or visit www.oracle.com if you need additional information or have any
20 * questions.
21 *
22 */
23
24 #include "precompiled.hpp"
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shenandoah/shenandoahForwarding.hpp"
27 #include "gc/shenandoah/shenandoahHeap.hpp"
28 #include "gc/shenandoah/shenandoahHeuristics.hpp"
29 #include "gc/shenandoah/shenandoahRuntime.hpp"
30 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
31 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
32 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/escape.hpp"
35 #include "opto/graphKit.hpp"
36 #include "opto/idealKit.hpp"
37 #include "opto/macro.hpp"
38 #include "opto/movenode.hpp"
39 #include "opto/narrowptrnode.hpp"
40 #include "opto/rootnode.hpp"
41
42 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() {
43 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2());
44 }
45
46 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena)
47 : _enqueue_barriers(new (comp_arena) GrowableArray<ShenandoahEnqueueBarrierNode*>(comp_arena, 8, 0, NULL)),
48 _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8, 0, NULL)) {
49 }
50
51 int ShenandoahBarrierSetC2State::enqueue_barriers_count() const {
52 return _enqueue_barriers->length();
53 }
54
55 ShenandoahEnqueueBarrierNode* ShenandoahBarrierSetC2State::enqueue_barrier(int idx) const {
56 return _enqueue_barriers->at(idx);
57 }
58
59 void ShenandoahBarrierSetC2State::add_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) {
60 assert(!_enqueue_barriers->contains(n), "duplicate entry in barrier list");
61 _enqueue_barriers->append(n);
62 }
63
64 void ShenandoahBarrierSetC2State::remove_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) {
65 if (_enqueue_barriers->contains(n)) {
66 _enqueue_barriers->remove(n);
67 }
68 }
69
70 int ShenandoahBarrierSetC2State::load_reference_barriers_count() const {
71 return _load_reference_barriers->length();
72 }
73
74 ShenandoahLoadReferenceBarrierNode* ShenandoahBarrierSetC2State::load_reference_barrier(int idx) const {
75 return _load_reference_barriers->at(idx);
76 }
77
78 void ShenandoahBarrierSetC2State::add_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
79 assert(!_load_reference_barriers->contains(n), "duplicate entry in barrier list");
80 _load_reference_barriers->append(n);
81 }
82
83 void ShenandoahBarrierSetC2State::remove_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
84 if (_load_reference_barriers->contains(n)) {
85 _load_reference_barriers->remove(n);
86 }
87 }
88
89 Node* ShenandoahBarrierSetC2::shenandoah_storeval_barrier(GraphKit* kit, Node* obj) const {
90 if (ShenandoahStoreValEnqueueBarrier) {
91 obj = shenandoah_enqueue_barrier(kit, obj);
92 }
93 return obj;
94 }
95
96 #define __ kit->
97
98 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseTransform* phase, Node* adr,
99 BasicType bt, uint adr_idx) const {
100 intptr_t offset = 0;
101 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
102 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
103
104 if (offset == Type::OffsetBot) {
105 return false; // cannot unalias unless there are precise offsets
106 }
107
108 if (alloc == NULL) {
109 return false; // No allocation found
110 }
111
112 intptr_t size_in_bytes = type2aelembytes(bt);
113
114 Node* mem = __ memory(adr_idx); // start searching here...
115
116 for (int cnt = 0; cnt < 50; cnt++) {
117
118 if (mem->is_Store()) {
119
120 Node* st_adr = mem->in(MemNode::Address);
121 intptr_t st_offset = 0;
122 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
123
124 if (st_base == NULL) {
125 break; // inscrutable pointer
126 }
127
128 // Break we have found a store with same base and offset as ours so break
129 if (st_base == base && st_offset == offset) {
130 break;
131 }
132
133 if (st_offset != offset && st_offset != Type::OffsetBot) {
134 const int MAX_STORE = BytesPerLong;
135 if (st_offset >= offset + size_in_bytes ||
136 st_offset <= offset - MAX_STORE ||
137 st_offset <= offset - mem->as_Store()->memory_size()) {
138 // Success: The offsets are provably independent.
139 // (You may ask, why not just test st_offset != offset and be done?
140 // The answer is that stores of different sizes can co-exist
141 // in the same sequence of RawMem effects. We sometimes initialize
142 // a whole 'tile' of array elements with a single jint or jlong.)
143 mem = mem->in(MemNode::Memory);
144 continue; // advance through independent store memory
145 }
146 }
147
148 if (st_base != base
149 && MemNode::detect_ptr_independence(base, alloc, st_base,
150 AllocateNode::Ideal_allocation(st_base, phase),
151 phase)) {
152 // Success: The bases are provably independent.
153 mem = mem->in(MemNode::Memory);
154 continue; // advance through independent store memory
155 }
156 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
157
158 InitializeNode* st_init = mem->in(0)->as_Initialize();
159 AllocateNode* st_alloc = st_init->allocation();
160
161 // Make sure that we are looking at the same allocation site.
162 // The alloc variable is guaranteed to not be null here from earlier check.
163 if (alloc == st_alloc) {
164 // Check that the initialization is storing NULL so that no previous store
165 // has been moved up and directly write a reference
166 Node* captured_store = st_init->find_captured_store(offset,
167 type2aelembytes(T_OBJECT),
168 phase);
169 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
170 return true;
171 }
172 }
173 }
174
175 // Unless there is an explicit 'continue', we must bail out here,
176 // because 'mem' is an inscrutable memory state (e.g., a call).
177 break;
178 }
179
180 return false;
181 }
182
183 #undef __
184 #define __ ideal.
185
186 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit,
187 bool do_load,
188 Node* obj,
189 Node* adr,
190 uint alias_idx,
191 Node* val,
192 const TypeOopPtr* val_type,
193 Node* pre_val,
194 BasicType bt) const {
195 // Some sanity checks
196 // Note: val is unused in this routine.
197
198 if (do_load) {
199 // We need to generate the load of the previous value
200 assert(obj != NULL, "must have a base");
201 assert(adr != NULL, "where are loading from?");
202 assert(pre_val == NULL, "loaded already?");
203 assert(val_type != NULL, "need a type");
204
205 if (ReduceInitialCardMarks
206 && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
207 return;
208 }
209
210 } else {
211 // In this case both val_type and alias_idx are unused.
212 assert(pre_val != NULL, "must be loaded already");
213 // Nothing to be done if pre_val is null.
214 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
215 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
216 }
217 assert(bt == T_OBJECT, "or we shouldn't be here");
218
219 IdealKit ideal(kit, true);
220
221 Node* tls = __ thread(); // ThreadLocalStorage
222
223 Node* no_base = __ top();
224 Node* zero = __ ConI(0);
225 Node* zeroX = __ ConX(0);
226
227 float likely = PROB_LIKELY(0.999);
228 float unlikely = PROB_UNLIKELY(0.999);
229
230 // Offsets into the thread
231 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
232 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
233
234 // Now the actual pointers into the thread
235 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
236 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
237
238 // Now some of the values
239 Node* marking;
240 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())));
241 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
242 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
243 assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape");
244
245 // if (!marking)
246 __ if_then(marking, BoolTest::ne, zero, unlikely); {
247 BasicType index_bt = TypeX_X->basic_type();
248 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
249 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
250
251 if (do_load) {
252 // load original value
253 // alias_idx correct??
254 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
255 }
256
257 // if (pre_val != NULL)
258 __ if_then(pre_val, BoolTest::ne, kit->null()); {
259 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
260
261 // is the queue for this thread full?
262 __ if_then(index, BoolTest::ne, zeroX, likely); {
263
264 // decrement the index
265 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
266
267 // Now get the buffer location we will log the previous value into and store it
268 Node *log_addr = __ AddP(no_base, buffer, next_index);
269 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
270 // update the index
271 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
272
273 } __ else_(); {
274
275 // logging buffer is full, call the runtime
276 const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type();
277 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls);
278 } __ end_if(); // (!index)
279 } __ end_if(); // (pre_val != NULL)
280 } __ end_if(); // (!marking)
281
282 // Final sync IdealKit and GraphKit.
283 kit->final_sync(ideal);
284
285 if (ShenandoahSATBBarrier && adr != NULL) {
286 Node* c = kit->control();
287 Node* call = c->in(1)->in(1)->in(1)->in(0);
288 assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected");
289 call->add_req(adr);
290 }
291 }
292
293 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) {
294 return call->is_CallLeaf() &&
295 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry);
296 }
297
298 bool ShenandoahBarrierSetC2::is_shenandoah_wb_call(Node* call) {
299 return call->is_CallLeaf() &&
300 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_JRT);
301 }
302
303 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseTransform *phase, Node* n) {
304 if (n->Opcode() != Op_If) {
305 return false;
306 }
307
308 Node* bol = n->in(1);
309 assert(bol->is_Bool(), "");
310 Node* cmpx = bol->in(1);
311 if (bol->as_Bool()->_test._test == BoolTest::ne &&
312 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
313 is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
314 cmpx->in(1)->in(2)->is_Con() &&
315 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
316 return true;
317 }
318
319 return false;
320 }
321
322 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
323 if (!n->is_Load()) return false;
324 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
325 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
326 && n->in(2)->in(3)->is_Con()
327 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
328 }
329
330 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
331 bool do_load,
332 Node* obj,
333 Node* adr,
334 uint alias_idx,
335 Node* val,
336 const TypeOopPtr* val_type,
337 Node* pre_val,
338 BasicType bt) const {
339 if (ShenandoahSATBBarrier) {
340 IdealKit ideal(kit);
341 kit->sync_kit(ideal);
342
343 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
344
345 ideal.sync_kit(kit);
346 kit->final_sync(ideal);
347 }
348 }
349
350 Node* ShenandoahBarrierSetC2::shenandoah_enqueue_barrier(GraphKit* kit, Node* pre_val) const {
351 return kit->gvn().transform(new ShenandoahEnqueueBarrierNode(pre_val));
352 }
353
354 // Helper that guards and inserts a pre-barrier.
355 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
356 Node* pre_val, bool need_mem_bar) const {
357 // We could be accessing the referent field of a reference object. If so, when G1
358 // is enabled, we need to log the value in the referent field in an SATB buffer.
359 // This routine performs some compile time filters and generates suitable
360 // runtime filters that guard the pre-barrier code.
361 // Also add memory barrier for non volatile load from the referent field
362 // to prevent commoning of loads across safepoint.
363
364 // Some compile time checks.
365
366 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
367 const TypeX* otype = offset->find_intptr_t_type();
368 if (otype != NULL && otype->is_con() &&
369 otype->get_con() != java_lang_ref_Reference::referent_offset) {
370 // Constant offset but not the reference_offset so just return
371 return;
372 }
373
374 // We only need to generate the runtime guards for instances.
375 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
376 if (btype != NULL) {
377 if (btype->isa_aryptr()) {
378 // Array type so nothing to do
379 return;
380 }
381
382 const TypeInstPtr* itype = btype->isa_instptr();
383 if (itype != NULL) {
384 // Can the klass of base_oop be statically determined to be
385 // _not_ a sub-class of Reference and _not_ Object?
386 ciKlass* klass = itype->klass();
387 if ( klass->is_loaded() &&
388 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
389 !kit->env()->Object_klass()->is_subtype_of(klass)) {
390 return;
391 }
392 }
393 }
394
395 // The compile time filters did not reject base_oop/offset so
396 // we need to generate the following runtime filters
397 //
398 // if (offset == java_lang_ref_Reference::_reference_offset) {
399 // if (instance_of(base, java.lang.ref.Reference)) {
400 // pre_barrier(_, pre_val, ...);
401 // }
402 // }
403
404 float likely = PROB_LIKELY( 0.999);
405 float unlikely = PROB_UNLIKELY(0.999);
406
407 IdealKit ideal(kit);
408
409 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
410
411 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
412 // Update graphKit memory and control from IdealKit.
413 kit->sync_kit(ideal);
414
415 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
416 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
417
418 // Update IdealKit memory and control from graphKit.
419 __ sync_kit(kit);
420
421 Node* one = __ ConI(1);
422 // is_instof == 0 if base_oop == NULL
423 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
424
425 // Update graphKit from IdeakKit.
426 kit->sync_kit(ideal);
427
428 // Use the pre-barrier to record the value in the referent field
429 satb_write_barrier_pre(kit, false /* do_load */,
430 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
431 pre_val /* pre_val */,
432 T_OBJECT);
433 if (need_mem_bar) {
434 // Add memory barrier to prevent commoning reads from this field
435 // across safepoint since GC can change its value.
436 kit->insert_mem_bar(Op_MemBarCPUOrder);
437 }
438 // Update IdealKit from graphKit.
439 __ sync_kit(kit);
440
441 } __ end_if(); // _ref_type != ref_none
442 } __ end_if(); // offset == referent_offset
443
444 // Final sync IdealKit and GraphKit.
445 kit->final_sync(ideal);
446 }
447
448 #undef __
449
450 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() {
451 const Type **fields = TypeTuple::fields(2);
452 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
453 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
454 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
455
456 // create result type (range)
457 fields = TypeTuple::fields(0);
458 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
459
460 return TypeFunc::make(domain, range);
461 }
462
463 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() {
464 const Type **fields = TypeTuple::fields(1);
465 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
466 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
467
468 // create result type (range)
469 fields = TypeTuple::fields(0);
470 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
471
472 return TypeFunc::make(domain, range);
473 }
474
475 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_write_barrier_Type() {
476 const Type **fields = TypeTuple::fields(1);
477 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
478 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
479
480 // create result type (range)
481 fields = TypeTuple::fields(1);
482 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL;
483 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
484
485 return TypeFunc::make(domain, range);
486 }
487
488 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
489 DecoratorSet decorators = access.decorators();
490
491 const TypePtr* adr_type = access.addr().type();
492 Node* adr = access.addr().node();
493
494 bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
495 bool on_heap = (decorators & IN_HEAP) != 0;
496
497 if (!access.is_oop() || (!on_heap && !anonymous)) {
498 return BarrierSetC2::store_at_resolved(access, val);
499 }
500
501 if (access.is_parse_access()) {
502 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
503 GraphKit* kit = parse_access.kit();
504
505 uint adr_idx = kit->C->get_alias_index(adr_type);
506 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
507 Node* value = val.node();
508 value = shenandoah_storeval_barrier(kit, value);
509 val.set_node(value);
510 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
511 static_cast<const TypeOopPtr*>(val.type()), NULL /* pre_val */, access.type());
512 } else {
513 assert(access.is_opt_access(), "only for optimization passes");
514 assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code");
515 C2OptAccess& opt_access = static_cast<C2OptAccess&>(access);
516 PhaseGVN& gvn = opt_access.gvn();
517 MergeMemNode* mm = opt_access.mem();
518
519 if (ShenandoahStoreValEnqueueBarrier) {
520 Node* enqueue = gvn.transform(new ShenandoahEnqueueBarrierNode(val.node()));
521 val.set_node(enqueue);
522 }
523 }
524 return BarrierSetC2::store_at_resolved(access, val);
525 }
526
527 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
528 DecoratorSet decorators = access.decorators();
529
530 Node* adr = access.addr().node();
531 Node* obj = access.base();
532
533 bool mismatched = (decorators & C2_MISMATCHED) != 0;
534 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
535 bool on_heap = (decorators & IN_HEAP) != 0;
536 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
537 bool is_unordered = (decorators & MO_UNORDERED) != 0;
538 bool need_cpu_mem_bar = !is_unordered || mismatched || !on_heap;
539
540 Node* top = Compile::current()->top();
541
542 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
543 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
544
545 if (access.is_oop()) {
546 if (ShenandoahLoadRefBarrier) {
547 load = new ShenandoahLoadReferenceBarrierNode(NULL, load);
548 if (access.is_parse_access()) {
549 load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load);
550 } else {
551 load = static_cast<C2OptAccess &>(access).gvn().transform(load);
552 }
553 }
554 }
555
556 // If we are reading the value of the referent field of a Reference
557 // object (either by using Unsafe directly or through reflection)
558 // then, if SATB is enabled, we need to record the referent in an
559 // SATB log buffer using the pre-barrier mechanism.
560 // Also we need to add memory barrier to prevent commoning reads
561 // from this field across safepoint since GC can change its value.
562 bool need_read_barrier = ShenandoahKeepAliveBarrier &&
563 (on_heap && (on_weak || (unknown && offset != top && obj != top)));
564
565 if (!access.is_oop() || !need_read_barrier) {
566 return load;
567 }
568
569 assert(access.is_parse_access(), "entry not supported at optimization time");
570 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
571 GraphKit* kit = parse_access.kit();
572
573 if (on_weak) {
574 // Use the pre-barrier to record the value in the referent field
575 satb_write_barrier_pre(kit, false /* do_load */,
576 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
577 load /* pre_val */, T_OBJECT);
578 // Add memory barrier to prevent commoning reads from this field
579 // across safepoint since GC can change its value.
580 kit->insert_mem_bar(Op_MemBarCPUOrder);
581 } else if (unknown) {
582 // We do not require a mem bar inside pre_barrier if need_mem_bar
583 // is set: the barriers would be emitted by us.
584 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
585 }
586
587 return load;
588 }
589
590 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
591 Node* new_val, const Type* value_type) const {
592 GraphKit* kit = access.kit();
593 if (access.is_oop()) {
594 new_val = shenandoah_storeval_barrier(kit, new_val);
595 shenandoah_write_barrier_pre(kit, false /* do_load */,
596 NULL, NULL, max_juint, NULL, NULL,
597 expected_val /* pre_val */, T_OBJECT);
598
599 MemNode::MemOrd mo = access.mem_node_mo();
600 Node* mem = access.memory();
601 Node* adr = access.addr().node();
602 const TypePtr* adr_type = access.addr().type();
603 Node* load_store = NULL;
604
605 #ifdef _LP64
606 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
607 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
608 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
609 if (ShenandoahCASBarrier) {
610 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
611 } else {
612 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
613 }
614 } else
615 #endif
616 {
617 if (ShenandoahCASBarrier) {
618 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
619 } else {
620 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
621 }
622 }
623
624 access.set_raw_access(load_store);
625 pin_atomic_op(access);
626
627 #ifdef _LP64
628 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
629 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
630 }
631 #endif
632 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, load_store));
633 return load_store;
634 }
635 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
636 }
637
638 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
639 Node* new_val, const Type* value_type) const {
640 GraphKit* kit = access.kit();
641 if (access.is_oop()) {
642 new_val = shenandoah_storeval_barrier(kit, new_val);
643 shenandoah_write_barrier_pre(kit, false /* do_load */,
644 NULL, NULL, max_juint, NULL, NULL,
645 expected_val /* pre_val */, T_OBJECT);
646 DecoratorSet decorators = access.decorators();
647 MemNode::MemOrd mo = access.mem_node_mo();
648 Node* mem = access.memory();
649 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
650 Node* load_store = NULL;
651 Node* adr = access.addr().node();
652 #ifdef _LP64
653 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
654 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
655 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
656 if (ShenandoahCASBarrier) {
657 if (is_weak_cas) {
658 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
659 } else {
660 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
661 }
662 } else {
663 if (is_weak_cas) {
664 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
665 } else {
666 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
667 }
668 }
669 } else
670 #endif
671 {
672 if (ShenandoahCASBarrier) {
673 if (is_weak_cas) {
674 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
675 } else {
676 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
677 }
678 } else {
679 if (is_weak_cas) {
680 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
681 } else {
682 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
683 }
684 }
685 }
686 access.set_raw_access(load_store);
687 pin_atomic_op(access);
688 return load_store;
689 }
690 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
691 }
692
693 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
694 GraphKit* kit = access.kit();
695 if (access.is_oop()) {
696 val = shenandoah_storeval_barrier(kit, val);
697 }
698 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
699 if (access.is_oop()) {
700 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, result));
701 shenandoah_write_barrier_pre(kit, false /* do_load */,
702 NULL, NULL, max_juint, NULL, NULL,
703 result /* pre_val */, T_OBJECT);
704 }
705 return result;
706 }
707
708 void ShenandoahBarrierSetC2::clone(GraphKit* kit, Node* src, Node* dst, Node* size, bool is_array) const {
709 assert(!src->is_AddP(), "unexpected input");
710 BarrierSetC2::clone(kit, src, dst, size, is_array);
711 }
712
713 Node* ShenandoahBarrierSetC2::obj_allocate(PhaseMacroExpand* macro, Node* ctrl, Node* mem, Node* toobig_false, Node* size_in_bytes,
714 Node*& i_o, Node*& needgc_ctrl,
715 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem,
716 intx prefetch_lines) const {
717 PhaseIterGVN& igvn = macro->igvn();
718
719 // Allocate several words more for the Shenandoah brooks pointer.
720 size_in_bytes = new AddXNode(size_in_bytes, igvn.MakeConX(ShenandoahForwarding::byte_size()));
721 macro->transform_later(size_in_bytes);
722
723 Node* fast_oop = BarrierSetC2::obj_allocate(macro, ctrl, mem, toobig_false, size_in_bytes,
724 i_o, needgc_ctrl, fast_oop_ctrl, fast_oop_rawmem,
725 prefetch_lines);
726
727 // Bump up object for Shenandoah brooks pointer.
728 fast_oop = new AddPNode(macro->top(), fast_oop, igvn.MakeConX(ShenandoahForwarding::byte_size()));
729 macro->transform_later(fast_oop);
730
731 // Initialize Shenandoah brooks pointer to point to the object itself.
732 fast_oop_rawmem = macro->make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, ShenandoahForwarding::byte_offset(), fast_oop, T_OBJECT);
733
734 return fast_oop;
735 }
736
737 // Support for GC barriers emitted during parsing
738 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
739 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) return true;
740 if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) {
741 return false;
742 }
743 CallLeafNode *call = node->as_CallLeaf();
744 if (call->_name == NULL) {
745 return false;
746 }
747
748 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 ||
749 strcmp(call->_name, "shenandoah_cas_obj") == 0 ||
750 strcmp(call->_name, "shenandoah_wb_pre") == 0;
751 }
752
753 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
754 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
755 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
756 }
757 if (c->Opcode() == Op_ShenandoahEnqueueBarrier) {
758 c = c->in(1);
759 }
760 return c;
761 }
762
763 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
764 return !ShenandoahBarrierC2Support::expand(C, igvn);
765 }
766
767 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
768 if (mode == LoopOptsShenandoahExpand) {
769 assert(UseShenandoahGC, "only for shenandoah");
770 ShenandoahBarrierC2Support::pin_and_expand(phase);
771 return true;
772 } else if (mode == LoopOptsShenandoahPostExpand) {
773 assert(UseShenandoahGC, "only for shenandoah");
774 visited.Clear();
775 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
776 return true;
777 }
778 return false;
779 }
780
781 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, ArrayCopyPhase phase) const {
782 bool is_oop = type == T_OBJECT || type == T_ARRAY;
783 if (!is_oop) {
784 return false;
785 }
786 if (tightly_coupled_alloc) {
787 if (phase == Optimization) {
788 return false;
789 }
790 return !is_clone;
791 }
792 if (phase == Optimization) {
793 return !ShenandoahStoreValEnqueueBarrier;
794 }
795 return true;
796 }
797
798 bool ShenandoahBarrierSetC2::clone_needs_postbarrier(ArrayCopyNode *ac, PhaseIterGVN& igvn) {
799 Node* src = ac->in(ArrayCopyNode::Src);
800 const TypeOopPtr* src_type = igvn.type(src)->is_oopptr();
801 if (src_type->isa_instptr() != NULL) {
802 ciInstanceKlass* ik = src_type->klass()->as_instance_klass();
803 if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) {
804 if (ik->has_object_fields()) {
805 return true;
806 } else {
807 if (!src_type->klass_is_exact()) {
808 igvn.C->dependencies()->assert_leaf_type(ik);
809 }
810 }
811 } else {
812 return true;
813 }
814 } else if (src_type->isa_aryptr()) {
815 BasicType src_elem = src_type->klass()->as_array_klass()->element_type()->basic_type();
816 if (src_elem == T_OBJECT || src_elem == T_ARRAY) {
817 return true;
818 }
819 } else {
820 return true;
821 }
822 return false;
823 }
824
825 void ShenandoahBarrierSetC2::clone_barrier_at_expansion(ArrayCopyNode* ac, Node* call, PhaseIterGVN& igvn) const {
826 assert(ac->is_clonebasic(), "no other kind of arraycopy here");
827
828 if (!clone_needs_postbarrier(ac, igvn)) {
829 BarrierSetC2::clone_barrier_at_expansion(ac, call, igvn);
830 return;
831 }
832
833 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
834 Node* c = new ProjNode(call,TypeFunc::Control);
835 c = igvn.transform(c);
836 Node* m = new ProjNode(call, TypeFunc::Memory);
837 m = igvn.transform(m);
838
839 Node* dest = ac->in(ArrayCopyNode::Dest);
840 assert(dest->is_AddP(), "bad input");
841 Node* barrier_call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(),
842 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier),
843 "shenandoah_clone_barrier", raw_adr_type);
844 barrier_call->init_req(TypeFunc::Control, c);
845 barrier_call->init_req(TypeFunc::I_O , igvn.C->top());
846 barrier_call->init_req(TypeFunc::Memory , m);
847 barrier_call->init_req(TypeFunc::ReturnAdr, igvn.C->top());
848 barrier_call->init_req(TypeFunc::FramePtr, igvn.C->top());
849 barrier_call->init_req(TypeFunc::Parms+0, dest->in(AddPNode::Base));
850
851 barrier_call = igvn.transform(barrier_call);
852 c = new ProjNode(barrier_call,TypeFunc::Control);
853 c = igvn.transform(c);
854 m = new ProjNode(barrier_call, TypeFunc::Memory);
855 m = igvn.transform(m);
856
857 Node* out_c = ac->proj_out(TypeFunc::Control);
858 Node* out_m = ac->proj_out(TypeFunc::Memory);
859 igvn.replace_node(out_c, c);
860 igvn.replace_node(out_m, m);
861 }
862
863
864 // Support for macro expanded GC barriers
865 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
866 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
867 state()->add_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
868 }
869 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
870 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
871 }
872 }
873
874 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
875 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
876 state()->remove_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
877 }
878 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
879 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
880 }
881 }
882
883 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const {
884 if (is_shenandoah_wb_pre_call(n)) {
885 shenandoah_eliminate_wb_pre(n, ¯o->igvn());
886 }
887 }
888
889 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
890 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
891 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
892 c = c->unique_ctrl_out();
893 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
894 c = c->unique_ctrl_out();
895 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
896 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
897 assert(iff->is_If(), "expect test");
898 if (!is_shenandoah_marking_if(igvn, iff)) {
899 c = c->unique_ctrl_out();
900 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
901 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
902 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
903 }
904 Node* cmpx = iff->in(1)->in(1);
905 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
906 igvn->rehash_node_delayed(call);
907 call->del_req(call->req()-1);
908 }
909
910 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
911 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
912 igvn->add_users_to_worklist(node);
913 }
914 }
915
916 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
917 for (uint i = 0; i < useful.size(); i++) {
918 Node* n = useful.at(i);
919 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
920 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
921 C->record_for_igvn(n->fast_out(i));
922 }
923 }
924 }
925 for (int i = state()->enqueue_barriers_count() - 1; i >= 0; i--) {
926 ShenandoahEnqueueBarrierNode* n = state()->enqueue_barrier(i);
927 if (!useful.member(n)) {
928 state()->remove_enqueue_barrier(n);
929 }
930 }
931 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
932 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
933 if (!useful.member(n)) {
934 state()->remove_load_reference_barrier(n);
935 }
936 }
937 }
938
939 void ShenandoahBarrierSetC2::add_users_to_worklist(Unique_Node_List* worklist) const {}
940
941 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
942 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
943 }
944
945 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
946 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
947 }
948
949 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
950 // expanded later, then now is the time to do so.
951 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
952
953 #ifdef ASSERT
954 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
955 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeExpand) {
956 ShenandoahBarrierC2Support::verify(Compile::current()->root());
957 } else if (phase == BarrierSetC2::BeforeCodeGen) {
958 // Verify G1 pre-barriers
959 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset());
960
961 ResourceArea *area = Thread::current()->resource_area();
962 Unique_Node_List visited(area);
963 Node_List worklist(area);
964 // We're going to walk control flow backwards starting from the Root
965 worklist.push(compile->root());
966 while (worklist.size() > 0) {
967 Node *x = worklist.pop();
968 if (x == NULL || x == compile->top()) continue;
969 if (visited.member(x)) {
970 continue;
971 } else {
972 visited.push(x);
973 }
974
975 if (x->is_Region()) {
976 for (uint i = 1; i < x->req(); i++) {
977 worklist.push(x->in(i));
978 }
979 } else {
980 worklist.push(x->in(0));
981 // We are looking for the pattern:
982 // /->ThreadLocal
983 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
984 // \->ConI(0)
985 // We want to verify that the If and the LoadB have the same control
986 // See GraphKit::g1_write_barrier_pre()
987 if (x->is_If()) {
988 IfNode *iff = x->as_If();
989 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
990 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
991 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
992 && cmp->in(1)->is_Load()) {
993 LoadNode *load = cmp->in(1)->as_Load();
994 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
995 && load->in(2)->in(3)->is_Con()
996 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
997
998 Node *if_ctrl = iff->in(0);
999 Node *load_ctrl = load->in(0);
1000
1001 if (if_ctrl != load_ctrl) {
1002 // Skip possible CProj->NeverBranch in infinite loops
1003 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
1004 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
1005 if_ctrl = if_ctrl->in(0)->in(0);
1006 }
1007 }
1008 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
1009 }
1010 }
1011 }
1012 }
1013 }
1014 }
1015 }
1016 }
1017 #endif
1018
1019 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
1020 if (is_shenandoah_wb_pre_call(n)) {
1021 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1022 if (n->req() > cnt) {
1023 Node* addp = n->in(cnt);
1024 if (has_only_shenandoah_wb_pre_uses(addp)) {
1025 n->del_req(cnt);
1026 if (can_reshape) {
1027 phase->is_IterGVN()->_worklist.push(addp);
1028 }
1029 return n;
1030 }
1031 }
1032 }
1033 if (n->Opcode() == Op_CmpP) {
1034 Node* in1 = n->in(1);
1035 Node* in2 = n->in(2);
1036 if (in1->bottom_type() == TypePtr::NULL_PTR) {
1037 in2 = step_over_gc_barrier(in2);
1038 }
1039 if (in2->bottom_type() == TypePtr::NULL_PTR) {
1040 in1 = step_over_gc_barrier(in1);
1041 }
1042 PhaseIterGVN* igvn = phase->is_IterGVN();
1043 if (in1 != n->in(1)) {
1044 if (igvn != NULL) {
1045 n->set_req_X(1, in1, igvn);
1046 } else {
1047 n->set_req(1, in1);
1048 }
1049 assert(in2 == n->in(2), "only one change");
1050 return n;
1051 }
1052 if (in2 != n->in(2)) {
1053 if (igvn != NULL) {
1054 n->set_req_X(2, in2, igvn);
1055 } else {
1056 n->set_req(2, in2);
1057 }
1058 return n;
1059 }
1060 } else if (can_reshape &&
1061 n->Opcode() == Op_If &&
1062 ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1063 n->in(0) != NULL) {
1064 Node* dom = n->in(0);
1065 Node* prev_dom = n;
1066 int op = n->Opcode();
1067 int dist = 16;
1068 // Search up the dominator tree for another heap stable test
1069 while (dom->Opcode() != op || // Not same opcode?
1070 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1?
1071 prev_dom->in(0) != dom) { // One path of test does not dominate?
1072 if (dist < 0) return NULL;
1073
1074 dist--;
1075 prev_dom = dom;
1076 dom = IfNode::up_one_dom(dom);
1077 if (!dom) return NULL;
1078 }
1079
1080 // Check that we did not follow a loop back to ourselves
1081 if (n == dom) {
1082 return NULL;
1083 }
1084
1085 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN());
1086 }
1087
1088 return NULL;
1089 }
1090
1091 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1092 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1093 Node* u = n->fast_out(i);
1094 if (!is_shenandoah_wb_pre_call(u)) {
1095 return false;
1096 }
1097 }
1098 return n->outcnt() > 0;
1099 }
1100
1101 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode) const {
1102 switch (opcode) {
1103 case Op_CallLeaf:
1104 case Op_CallLeafNoFP: {
1105 assert (n->is_Call(), "");
1106 CallNode *call = n->as_Call();
1107 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1108 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1109 if (call->req() > cnt) {
1110 assert(call->req() == cnt + 1, "only one extra input");
1111 Node *addp = call->in(cnt);
1112 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1113 call->del_req(cnt);
1114 }
1115 }
1116 return false;
1117 }
1118 case Op_ShenandoahCompareAndSwapP:
1119 case Op_ShenandoahCompareAndSwapN:
1120 case Op_ShenandoahWeakCompareAndSwapN:
1121 case Op_ShenandoahWeakCompareAndSwapP:
1122 case Op_ShenandoahCompareAndExchangeP:
1123 case Op_ShenandoahCompareAndExchangeN:
1124 #ifdef ASSERT
1125 if( VerifyOptoOopOffsets ) {
1126 MemNode* mem = n->as_Mem();
1127 // Check to see if address types have grounded out somehow.
1128 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
1129 ciInstanceKlass *k = tp->klass()->as_instance_klass();
1130 bool oop_offset_is_sane = k->contains_field_offset(tp->offset());
1131 assert( !tp || oop_offset_is_sane, "" );
1132 }
1133 #endif
1134 return true;
1135 case Op_ShenandoahLoadReferenceBarrier:
1136 assert(false, "should have been expanded already");
1137 return true;
1138 default:
1139 return false;
1140 }
1141 }
1142
1143 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1144 switch (opcode) {
1145 case Op_ShenandoahCompareAndExchangeP:
1146 case Op_ShenandoahCompareAndExchangeN:
1147 conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1148 // fallthrough
1149 case Op_ShenandoahWeakCompareAndSwapP:
1150 case Op_ShenandoahWeakCompareAndSwapN:
1151 case Op_ShenandoahCompareAndSwapP:
1152 case Op_ShenandoahCompareAndSwapN:
1153 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1154 return true;
1155 case Op_StoreP: {
1156 Node* adr = n->in(MemNode::Address);
1157 const Type* adr_type = gvn->type(adr);
1158 // Pointer stores in G1 barriers looks like unsafe access.
1159 // Ignore such stores to be able scalar replace non-escaping
1160 // allocations.
1161 if (adr_type->isa_rawptr() && adr->is_AddP()) {
1162 Node* base = conn_graph->get_addp_base(adr);
1163 if (base->Opcode() == Op_LoadP &&
1164 base->in(MemNode::Address)->is_AddP()) {
1165 adr = base->in(MemNode::Address);
1166 Node* tls = conn_graph->get_addp_base(adr);
1167 if (tls->Opcode() == Op_ThreadLocal) {
1168 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1169 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1170 if (offs == buf_offset) {
1171 return true; // Pre barrier previous oop value store.
1172 }
1173 }
1174 }
1175 }
1176 return false;
1177 }
1178 case Op_ShenandoahEnqueueBarrier:
1179 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1180 break;
1181 case Op_ShenandoahLoadReferenceBarrier:
1182 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist);
1183 return true;
1184 default:
1185 // Nothing
1186 break;
1187 }
1188 return false;
1189 }
1190
1191 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1192 switch (opcode) {
1193 case Op_ShenandoahCompareAndExchangeP:
1194 case Op_ShenandoahCompareAndExchangeN: {
1195 Node *adr = n->in(MemNode::Address);
1196 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
1197 // fallthrough
1198 }
1199 case Op_ShenandoahCompareAndSwapP:
1200 case Op_ShenandoahCompareAndSwapN:
1201 case Op_ShenandoahWeakCompareAndSwapP:
1202 case Op_ShenandoahWeakCompareAndSwapN:
1203 return conn_graph->add_final_edges_unsafe_access(n, opcode);
1204 case Op_ShenandoahEnqueueBarrier:
1205 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), NULL);
1206 return true;
1207 case Op_ShenandoahLoadReferenceBarrier:
1208 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), NULL);
1209 return true;
1210 default:
1211 // Nothing
1212 break;
1213 }
1214 return false;
1215 }
1216
1217 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1218 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1219 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1220
1221 }
1222
1223 bool ShenandoahBarrierSetC2::escape_is_barrier_node(Node* n) const {
1224 return n->Opcode() == Op_ShenandoahLoadReferenceBarrier;
1225 }
1226
1227 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1228 switch (opcode) {
1229 case Op_ShenandoahCompareAndExchangeP:
1230 case Op_ShenandoahCompareAndExchangeN:
1231 case Op_ShenandoahWeakCompareAndSwapP:
1232 case Op_ShenandoahWeakCompareAndSwapN:
1233 case Op_ShenandoahCompareAndSwapP:
1234 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree
1235 Node* newval = n->in(MemNode::ValueIn);
1236 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1237 Node* pair = new BinaryNode(oldval, newval);
1238 n->set_req(MemNode::ValueIn,pair);
1239 n->del_req(LoadStoreConditionalNode::ExpectedIn);
1240 return true;
1241 }
1242 default:
1243 break;
1244 }
1245 return false;
1246 }
1247
1248 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1249 return xop == Op_ShenandoahCompareAndExchangeP ||
1250 xop == Op_ShenandoahCompareAndExchangeN ||
1251 xop == Op_ShenandoahWeakCompareAndSwapP ||
1252 xop == Op_ShenandoahWeakCompareAndSwapN ||
1253 xop == Op_ShenandoahCompareAndSwapN ||
1254 xop == Op_ShenandoahCompareAndSwapP;
1255 }