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/shenandoahConcurrentRoots.hpp"
27 #include "gc/shenandoah/shenandoahForwarding.hpp"
28 #include "gc/shenandoah/shenandoahHeap.hpp"
29 #include "gc/shenandoah/shenandoahHeuristics.hpp"
30 #include "gc/shenandoah/shenandoahRuntime.hpp"
31 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
32 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
33 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/escape.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/idealKit.hpp"
38 #include "opto/macro.hpp"
39 #include "opto/movenode.hpp"
40 #include "opto/narrowptrnode.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/runtime.hpp"
43
44 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() {
45 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2());
46 }
47
48 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena)
49 : _enqueue_barriers(new (comp_arena) GrowableArray<ShenandoahEnqueueBarrierNode*>(comp_arena, 8, 0, NULL)),
50 _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8, 0, NULL)) {
51 }
52
53 int ShenandoahBarrierSetC2State::enqueue_barriers_count() const {
54 return _enqueue_barriers->length();
55 }
56
57 ShenandoahEnqueueBarrierNode* ShenandoahBarrierSetC2State::enqueue_barrier(int idx) const {
58 return _enqueue_barriers->at(idx);
59 }
60
61 void ShenandoahBarrierSetC2State::add_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) {
62 assert(!_enqueue_barriers->contains(n), "duplicate entry in barrier list");
63 _enqueue_barriers->append(n);
64 }
65
66 void ShenandoahBarrierSetC2State::remove_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) {
67 if (_enqueue_barriers->contains(n)) {
68 _enqueue_barriers->remove(n);
69 }
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_storeval_barrier(GraphKit* kit, Node* obj) const {
92 if (ShenandoahStoreValEnqueueBarrier) {
93 obj = shenandoah_enqueue_barrier(kit, 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 G1 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)) ||
307 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_narrow)) ||
308 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_native));
309 }
310
311 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseTransform *phase, Node* n) {
312 if (n->Opcode() != Op_If) {
313 return false;
314 }
315
316 Node* bol = n->in(1);
317 assert(bol->is_Bool(), "");
318 Node* cmpx = bol->in(1);
319 if (bol->as_Bool()->_test._test == BoolTest::ne &&
320 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
321 is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
322 cmpx->in(1)->in(2)->is_Con() &&
323 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
324 return true;
325 }
326
327 return false;
328 }
329
330 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
331 if (!n->is_Load()) return false;
332 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
333 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
334 && n->in(2)->in(3)->is_Con()
335 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
336 }
337
338 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
339 bool do_load,
340 Node* obj,
341 Node* adr,
342 uint alias_idx,
343 Node* val,
344 const TypeOopPtr* val_type,
345 Node* pre_val,
346 BasicType bt) const {
347 if (ShenandoahSATBBarrier) {
348 IdealKit ideal(kit);
349 kit->sync_kit(ideal);
350
351 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
352
353 ideal.sync_kit(kit);
354 kit->final_sync(ideal);
355 }
356 }
357
358 Node* ShenandoahBarrierSetC2::shenandoah_enqueue_barrier(GraphKit* kit, Node* pre_val) const {
359 return kit->gvn().transform(new ShenandoahEnqueueBarrierNode(pre_val));
360 }
361
362 // Helper that guards and inserts a pre-barrier.
363 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
364 Node* pre_val, bool need_mem_bar) const {
365 // We could be accessing the referent field of a reference object. If so, when G1
366 // is enabled, we need to log the value in the referent field in an SATB buffer.
367 // This routine performs some compile time filters and generates suitable
368 // runtime filters that guard the pre-barrier code.
369 // Also add memory barrier for non volatile load from the referent field
370 // to prevent commoning of loads across safepoint.
371
372 // Some compile time checks.
373
374 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
375 const TypeX* otype = offset->find_intptr_t_type();
376 if (otype != NULL && otype->is_con() &&
377 otype->get_con() != java_lang_ref_Reference::referent_offset) {
378 // Constant offset but not the reference_offset so just return
379 return;
380 }
381
382 // We only need to generate the runtime guards for instances.
383 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
384 if (btype != NULL) {
385 if (btype->isa_aryptr()) {
386 // Array type so nothing to do
387 return;
388 }
389
390 const TypeInstPtr* itype = btype->isa_instptr();
391 if (itype != NULL) {
392 // Can the klass of base_oop be statically determined to be
393 // _not_ a sub-class of Reference and _not_ Object?
394 ciKlass* klass = itype->klass();
395 if ( klass->is_loaded() &&
396 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
397 !kit->env()->Object_klass()->is_subtype_of(klass)) {
398 return;
399 }
400 }
401 }
402
403 // The compile time filters did not reject base_oop/offset so
404 // we need to generate the following runtime filters
405 //
406 // if (offset == java_lang_ref_Reference::_reference_offset) {
407 // if (instance_of(base, java.lang.ref.Reference)) {
408 // pre_barrier(_, pre_val, ...);
409 // }
410 // }
411
412 float likely = PROB_LIKELY( 0.999);
413 float unlikely = PROB_UNLIKELY(0.999);
414
415 IdealKit ideal(kit);
416
417 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
418
419 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
420 // Update graphKit memory and control from IdealKit.
421 kit->sync_kit(ideal);
422
423 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
424 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
425
426 // Update IdealKit memory and control from graphKit.
427 __ sync_kit(kit);
428
429 Node* one = __ ConI(1);
430 // is_instof == 0 if base_oop == NULL
431 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
432
433 // Update graphKit from IdeakKit.
434 kit->sync_kit(ideal);
435
436 // Use the pre-barrier to record the value in the referent field
437 satb_write_barrier_pre(kit, false /* do_load */,
438 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
439 pre_val /* pre_val */,
440 T_OBJECT);
441 if (need_mem_bar) {
442 // Add memory barrier to prevent commoning reads from this field
443 // across safepoint since GC can change its value.
444 kit->insert_mem_bar(Op_MemBarCPUOrder);
445 }
446 // Update IdealKit from graphKit.
447 __ sync_kit(kit);
448
449 } __ end_if(); // _ref_type != ref_none
450 } __ end_if(); // offset == referent_offset
451
452 // Final sync IdealKit and GraphKit.
453 kit->final_sync(ideal);
454 }
455
456 #undef __
457
458 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() {
459 const Type **fields = TypeTuple::fields(2);
460 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
461 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
462 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
463
464 // create result type (range)
465 fields = TypeTuple::fields(0);
466 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
467
468 return TypeFunc::make(domain, range);
469 }
470
471 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() {
472 const Type **fields = TypeTuple::fields(1);
473 fields[TypeFunc::Parms+0] = TypeOopPtr::NOTNULL; // src oop
474 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
475
476 // create result type (range)
477 fields = TypeTuple::fields(0);
478 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
479
480 return TypeFunc::make(domain, range);
481 }
482
483 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type() {
484 const Type **fields = TypeTuple::fields(2);
485 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
486 fields[TypeFunc::Parms+1] = TypeRawPtr::BOTTOM; // original load address
487
488 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
489
490 // create result type (range)
491 fields = TypeTuple::fields(1);
492 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL;
493 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
494
495 return TypeFunc::make(domain, range);
496 }
497
498 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
499 DecoratorSet decorators = access.decorators();
500
501 const TypePtr* adr_type = access.addr().type();
502 Node* adr = access.addr().node();
503
504 bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
505 bool on_heap = (decorators & IN_HEAP) != 0;
506
507 if (!access.is_oop() || (!on_heap && !anonymous)) {
508 return BarrierSetC2::store_at_resolved(access, val);
509 }
510
511 if (access.is_parse_access()) {
512 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
513 GraphKit* kit = parse_access.kit();
514
515 uint adr_idx = kit->C->get_alias_index(adr_type);
516 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
517 Node* value = val.node();
518 value = shenandoah_storeval_barrier(kit, value);
519 val.set_node(value);
520 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
521 static_cast<const TypeOopPtr*>(val.type()), NULL /* pre_val */, access.type());
522 } else {
523 assert(access.is_opt_access(), "only for optimization passes");
524 assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code");
525 C2OptAccess& opt_access = static_cast<C2OptAccess&>(access);
526 PhaseGVN& gvn = opt_access.gvn();
527 MergeMemNode* mm = opt_access.mem();
528
529 if (ShenandoahStoreValEnqueueBarrier) {
530 Node* enqueue = gvn.transform(new ShenandoahEnqueueBarrierNode(val.node()));
531 val.set_node(enqueue);
532 }
533 }
534 return BarrierSetC2::store_at_resolved(access, val);
535 }
536
537 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
538 // 1: load reference
539 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
540 // For none-reference load, no additional barrier is needed
541 if (!access.is_oop()) {
542 return load;
543 }
544
545 DecoratorSet decorators = access.decorators();
546 bool in_native = (decorators & IN_NATIVE) != 0;
547
548 // 2: apply LRB if ShenandoahLoadRefBarrier is set
549 if (ShenandoahLoadRefBarrier) {
550 // Native barrier is for concurrent root processing
551 bool use_native_barrier = in_native && ShenandoahConcurrentRoots::can_do_concurrent_roots();
552 load = new ShenandoahLoadReferenceBarrierNode(NULL, load, use_native_barrier);
553 if (access.is_parse_access()) {
554 load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load);
555 } else {
556 load = static_cast<C2OptAccess &>(access).gvn().transform(load);
557 }
558 }
559
560 // 3: apply keep-alive barrier if ShenandoahKeepAliveBarrier is set
561 if (ShenandoahKeepAliveBarrier) {
562 Node* top = Compile::current()->top();
563 Node* adr = access.addr().node();
564 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
565 Node* obj = access.base();
566
567 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
568 bool on_weak_ref = (decorators & (ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF)) != 0;
569 bool is_traversal_mode = ShenandoahHeap::heap()->is_traversal_mode();
570 bool keep_alive = (decorators & AS_NO_KEEPALIVE) == 0 || is_traversal_mode;
571
572 // If we are reading the value of the referent field of a Reference
573 // object (either by using Unsafe directly or through reflection)
574 // then, if SATB is enabled, we need to record the referent in an
575 // SATB log buffer using the pre-barrier mechanism.
576 // Also we need to add memory barrier to prevent commoning reads
577 // from this field across safepoint since GC can change its value.
578 if (!on_weak_ref || (unknown && (offset == top || obj == top)) || !keep_alive) {
579 return load;
580 }
581
582 assert(access.is_parse_access(), "entry not supported at optimization time");
583 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
584 GraphKit* kit = parse_access.kit();
585 bool mismatched = (decorators & C2_MISMATCHED) != 0;
586 bool is_unordered = (decorators & MO_UNORDERED) != 0;
587 bool need_cpu_mem_bar = !is_unordered || mismatched || in_native;
588
589 if (on_weak_ref) {
590 // Use the pre-barrier to record the value in the referent field
591 satb_write_barrier_pre(kit, false /* do_load */,
592 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
593 load /* pre_val */, T_OBJECT);
594 // Add memory barrier to prevent commoning reads from this field
595 // across safepoint since GC can change its value.
596 kit->insert_mem_bar(Op_MemBarCPUOrder);
597 } else if (unknown) {
598 // We do not require a mem bar inside pre_barrier if need_mem_bar
599 // is set: the barriers would be emitted by us.
600 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
601 }
602 }
603
604 return load;
605 }
606
607 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
608 Node* new_val, const Type* value_type) const {
609 GraphKit* kit = access.kit();
610 if (access.is_oop()) {
611 new_val = shenandoah_storeval_barrier(kit, new_val);
612 shenandoah_write_barrier_pre(kit, false /* do_load */,
613 NULL, NULL, max_juint, NULL, NULL,
614 expected_val /* pre_val */, T_OBJECT);
615
616 MemNode::MemOrd mo = access.mem_node_mo();
617 Node* mem = access.memory();
618 Node* adr = access.addr().node();
619 const TypePtr* adr_type = access.addr().type();
620 Node* load_store = NULL;
621
622 #ifdef _LP64
623 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
624 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
625 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
626 if (ShenandoahCASBarrier) {
627 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
628 } else {
629 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
630 }
631 } else
632 #endif
633 {
634 if (ShenandoahCASBarrier) {
635 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
636 } else {
637 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
638 }
639 }
640
641 access.set_raw_access(load_store);
642 pin_atomic_op(access);
643
644 #ifdef _LP64
645 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
646 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
647 }
648 #endif
649 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, load_store, false));
650 return load_store;
651 }
652 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
653 }
654
655 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
656 Node* new_val, const Type* value_type) const {
657 GraphKit* kit = access.kit();
658 if (access.is_oop()) {
659 new_val = shenandoah_storeval_barrier(kit, new_val);
660 shenandoah_write_barrier_pre(kit, false /* do_load */,
661 NULL, NULL, max_juint, NULL, NULL,
662 expected_val /* pre_val */, T_OBJECT);
663 DecoratorSet decorators = access.decorators();
664 MemNode::MemOrd mo = access.mem_node_mo();
665 Node* mem = access.memory();
666 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
667 Node* load_store = NULL;
668 Node* adr = access.addr().node();
669 #ifdef _LP64
670 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
671 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
672 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
673 if (ShenandoahCASBarrier) {
674 if (is_weak_cas) {
675 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
676 } else {
677 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
678 }
679 } else {
680 if (is_weak_cas) {
681 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
682 } else {
683 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
684 }
685 }
686 } else
687 #endif
688 {
689 if (ShenandoahCASBarrier) {
690 if (is_weak_cas) {
691 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
692 } else {
693 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
694 }
695 } else {
696 if (is_weak_cas) {
697 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
698 } else {
699 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
700 }
701 }
702 }
703 access.set_raw_access(load_store);
704 pin_atomic_op(access);
705 return load_store;
706 }
707 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
708 }
709
710 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
711 GraphKit* kit = access.kit();
712 if (access.is_oop()) {
713 val = shenandoah_storeval_barrier(kit, val);
714 }
715 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
716 if (access.is_oop()) {
717 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, result, false));
718 shenandoah_write_barrier_pre(kit, false /* do_load */,
719 NULL, NULL, max_juint, NULL, NULL,
720 result /* pre_val */, T_OBJECT);
721 }
722 return result;
723 }
724
725 // Support for GC barriers emitted during parsing
726 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
727 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) return true;
728 if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) {
729 return false;
730 }
731 CallLeafNode *call = node->as_CallLeaf();
732 if (call->_name == NULL) {
733 return false;
734 }
735
736 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 ||
737 strcmp(call->_name, "shenandoah_cas_obj") == 0 ||
738 strcmp(call->_name, "shenandoah_wb_pre") == 0;
739 }
740
741 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
742 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
743 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
744 }
745 if (c->Opcode() == Op_ShenandoahEnqueueBarrier) {
746 c = c->in(1);
747 }
748 return c;
749 }
750
751 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
752 return !ShenandoahBarrierC2Support::expand(C, igvn);
753 }
754
755 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
756 if (mode == LoopOptsShenandoahExpand) {
757 assert(UseShenandoahGC, "only for shenandoah");
758 ShenandoahBarrierC2Support::pin_and_expand(phase);
759 return true;
760 } else if (mode == LoopOptsShenandoahPostExpand) {
761 assert(UseShenandoahGC, "only for shenandoah");
762 visited.Clear();
763 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
764 return true;
765 }
766 return false;
767 }
768
769 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, ArrayCopyPhase phase) const {
770 bool is_oop = is_reference_type(type);
771 if (!is_oop) {
772 return false;
773 }
774 if (tightly_coupled_alloc) {
775 if (phase == Optimization) {
776 return false;
777 }
778 return !is_clone;
779 }
780 if (phase == Optimization) {
781 return !ShenandoahStoreValEnqueueBarrier;
782 }
783 return true;
784 }
785
786 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
787 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
788 if (src_type->isa_instptr() != NULL) {
789 ciInstanceKlass* ik = src_type->klass()->as_instance_klass();
790 if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) {
791 if (ik->has_object_fields()) {
792 return true;
793 } else {
794 if (!src_type->klass_is_exact()) {
795 Compile::current()->dependencies()->assert_leaf_type(ik);
796 }
797 }
798 } else {
799 return true;
800 }
801 } else if (src_type->isa_aryptr()) {
802 BasicType src_elem = src_type->klass()->as_array_klass()->element_type()->basic_type();
803 if (is_reference_type(src_elem)) {
804 return true;
805 }
806 } else {
807 return true;
808 }
809 return false;
810 }
811
812 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
813 Node* ctrl = ac->in(TypeFunc::Control);
814 Node* mem = ac->in(TypeFunc::Memory);
815 Node* src = ac->in(ArrayCopyNode::Src);
816 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
817 Node* dest = ac->in(ArrayCopyNode::Dest);
818 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
819 Node* length = ac->in(ArrayCopyNode::Length);
820 assert (src_offset == NULL && dest_offset == NULL, "for clone offsets should be null");
821 assert (src->is_AddP(), "for clone the src should be the interior ptr");
822 assert (dest->is_AddP(), "for clone the dst should be the interior ptr");
823
824 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
825 // Check if heap is has forwarded objects. If it does, we need to call into the special
826 // routine that would fix up source references before we can continue.
827
828 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
829 Node* region = new RegionNode(PATH_LIMIT);
830 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM);
831
832 Node* thread = phase->transform_later(new ThreadLocalNode());
833 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
834 Node* gc_state_addr = phase->transform_later(new AddPNode(phase->C->top(), thread, offset));
835
836 uint gc_state_idx = Compile::AliasIdxRaw;
837 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument
838 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
839
840 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered));
841 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED)));
842 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT)));
843 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne));
844
845 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If();
846 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff));
847 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff));
848
849 // Heap is stable, no need to do anything additional
850 region->init_req(_heap_stable, stable_ctrl);
851 mem_phi->init_req(_heap_stable, mem);
852
853 // Heap is unstable, call into clone barrier stub
854 Node* call = phase->make_leaf_call(unstable_ctrl, mem,
855 ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(),
856 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier),
857 "shenandoah_clone",
858 TypeRawPtr::BOTTOM,
859 src->in(AddPNode::Base));
860 call = phase->transform_later(call);
861
862 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control));
863 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory));
864 region->init_req(_heap_unstable, ctrl);
865 mem_phi->init_req(_heap_unstable, mem);
866
867 // Wire up the actual arraycopy stub now
868 ctrl = phase->transform_later(region);
869 mem = phase->transform_later(mem_phi);
870
871 const char* name = "arraycopy";
872 call = phase->make_leaf_call(ctrl, mem,
873 OptoRuntime::fast_arraycopy_Type(),
874 phase->basictype2arraycopy(T_LONG, NULL, NULL, true, name, true),
875 name, TypeRawPtr::BOTTOM,
876 src, dest, length
877 LP64_ONLY(COMMA phase->top()));
878 call = phase->transform_later(call);
879
880 // Hook up the whole thing into the graph
881 phase->igvn().replace_node(ac, call);
882 } else {
883 BarrierSetC2::clone_at_expansion(phase, ac);
884 }
885 }
886
887
888 // Support for macro expanded GC barriers
889 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
890 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
891 state()->add_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
892 }
893 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
894 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
895 }
896 }
897
898 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
899 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
900 state()->remove_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
901 }
902 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
903 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
904 }
905 }
906
907 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const {
908 if (is_shenandoah_wb_pre_call(n)) {
909 shenandoah_eliminate_wb_pre(n, ¯o->igvn());
910 }
911 }
912
913 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
914 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
915 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
916 c = c->unique_ctrl_out();
917 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
918 c = c->unique_ctrl_out();
919 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
920 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
921 assert(iff->is_If(), "expect test");
922 if (!is_shenandoah_marking_if(igvn, iff)) {
923 c = c->unique_ctrl_out();
924 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
925 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
926 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
927 }
928 Node* cmpx = iff->in(1)->in(1);
929 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
930 igvn->rehash_node_delayed(call);
931 call->del_req(call->req()-1);
932 }
933
934 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
935 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
936 igvn->add_users_to_worklist(node);
937 }
938 }
939
940 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
941 for (uint i = 0; i < useful.size(); i++) {
942 Node* n = useful.at(i);
943 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
944 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
945 C->record_for_igvn(n->fast_out(i));
946 }
947 }
948 }
949 for (int i = state()->enqueue_barriers_count() - 1; i >= 0; i--) {
950 ShenandoahEnqueueBarrierNode* n = state()->enqueue_barrier(i);
951 if (!useful.member(n)) {
952 state()->remove_enqueue_barrier(n);
953 }
954 }
955 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
956 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
957 if (!useful.member(n)) {
958 state()->remove_load_reference_barrier(n);
959 }
960 }
961 }
962
963 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
964 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
965 }
966
967 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
968 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
969 }
970
971 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
972 // expanded later, then now is the time to do so.
973 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
974
975 #ifdef ASSERT
976 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
977 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) {
978 ShenandoahBarrierC2Support::verify(Compile::current()->root());
979 } else if (phase == BarrierSetC2::BeforeCodeGen) {
980 // Verify G1 pre-barriers
981 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset());
982
983 ResourceArea *area = Thread::current()->resource_area();
984 Unique_Node_List visited(area);
985 Node_List worklist(area);
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 if (in1->bottom_type() == TypePtr::NULL_PTR) {
1059 in2 = step_over_gc_barrier(in2);
1060 }
1061 if (in2->bottom_type() == TypePtr::NULL_PTR) {
1062 in1 = step_over_gc_barrier(in1);
1063 }
1064 PhaseIterGVN* igvn = phase->is_IterGVN();
1065 if (in1 != n->in(1)) {
1066 if (igvn != NULL) {
1067 n->set_req_X(1, in1, igvn);
1068 } else {
1069 n->set_req(1, in1);
1070 }
1071 assert(in2 == n->in(2), "only one change");
1072 return n;
1073 }
1074 if (in2 != n->in(2)) {
1075 if (igvn != NULL) {
1076 n->set_req_X(2, in2, igvn);
1077 } else {
1078 n->set_req(2, in2);
1079 }
1080 return n;
1081 }
1082 } else if (can_reshape &&
1083 n->Opcode() == Op_If &&
1084 ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1085 n->in(0) != NULL) {
1086 Node* dom = n->in(0);
1087 Node* prev_dom = n;
1088 int op = n->Opcode();
1089 int dist = 16;
1090 // Search up the dominator tree for another heap stable test
1091 while (dom->Opcode() != op || // Not same opcode?
1092 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1?
1093 prev_dom->in(0) != dom) { // One path of test does not dominate?
1094 if (dist < 0) return NULL;
1095
1096 dist--;
1097 prev_dom = dom;
1098 dom = IfNode::up_one_dom(dom);
1099 if (!dom) return NULL;
1100 }
1101
1102 // Check that we did not follow a loop back to ourselves
1103 if (n == dom) {
1104 return NULL;
1105 }
1106
1107 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN());
1108 }
1109
1110 return NULL;
1111 }
1112
1113 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1114 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1115 Node* u = n->fast_out(i);
1116 if (!is_shenandoah_wb_pre_call(u)) {
1117 return false;
1118 }
1119 }
1120 return n->outcnt() > 0;
1121 }
1122
1123 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode) const {
1124 switch (opcode) {
1125 case Op_CallLeaf:
1126 case Op_CallLeafNoFP: {
1127 assert (n->is_Call(), "");
1128 CallNode *call = n->as_Call();
1129 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1130 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1131 if (call->req() > cnt) {
1132 assert(call->req() == cnt + 1, "only one extra input");
1133 Node *addp = call->in(cnt);
1134 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1135 call->del_req(cnt);
1136 }
1137 }
1138 return false;
1139 }
1140 case Op_ShenandoahCompareAndSwapP:
1141 case Op_ShenandoahCompareAndSwapN:
1142 case Op_ShenandoahWeakCompareAndSwapN:
1143 case Op_ShenandoahWeakCompareAndSwapP:
1144 case Op_ShenandoahCompareAndExchangeP:
1145 case Op_ShenandoahCompareAndExchangeN:
1146 #ifdef ASSERT
1147 if( VerifyOptoOopOffsets ) {
1148 MemNode* mem = n->as_Mem();
1149 // Check to see if address types have grounded out somehow.
1150 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
1151 ciInstanceKlass *k = tp->klass()->as_instance_klass();
1152 bool oop_offset_is_sane = k->contains_field_offset(tp->offset());
1153 assert( !tp || oop_offset_is_sane, "" );
1154 }
1155 #endif
1156 return true;
1157 case Op_ShenandoahLoadReferenceBarrier:
1158 assert(false, "should have been expanded already");
1159 return true;
1160 default:
1161 return false;
1162 }
1163 }
1164
1165 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1166 switch (opcode) {
1167 case Op_ShenandoahCompareAndExchangeP:
1168 case Op_ShenandoahCompareAndExchangeN:
1169 conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1170 // fallthrough
1171 case Op_ShenandoahWeakCompareAndSwapP:
1172 case Op_ShenandoahWeakCompareAndSwapN:
1173 case Op_ShenandoahCompareAndSwapP:
1174 case Op_ShenandoahCompareAndSwapN:
1175 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1176 return true;
1177 case Op_StoreP: {
1178 Node* adr = n->in(MemNode::Address);
1179 const Type* adr_type = gvn->type(adr);
1180 // Pointer stores in G1 barriers looks like unsafe access.
1181 // Ignore such stores to be able scalar replace non-escaping
1182 // allocations.
1183 if (adr_type->isa_rawptr() && adr->is_AddP()) {
1184 Node* base = conn_graph->get_addp_base(adr);
1185 if (base->Opcode() == Op_LoadP &&
1186 base->in(MemNode::Address)->is_AddP()) {
1187 adr = base->in(MemNode::Address);
1188 Node* tls = conn_graph->get_addp_base(adr);
1189 if (tls->Opcode() == Op_ThreadLocal) {
1190 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1191 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1192 if (offs == buf_offset) {
1193 return true; // Pre barrier previous oop value store.
1194 }
1195 }
1196 }
1197 }
1198 return false;
1199 }
1200 case Op_ShenandoahEnqueueBarrier:
1201 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1202 break;
1203 case Op_ShenandoahLoadReferenceBarrier:
1204 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist);
1205 return true;
1206 default:
1207 // Nothing
1208 break;
1209 }
1210 return false;
1211 }
1212
1213 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1214 switch (opcode) {
1215 case Op_ShenandoahCompareAndExchangeP:
1216 case Op_ShenandoahCompareAndExchangeN: {
1217 Node *adr = n->in(MemNode::Address);
1218 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
1219 // fallthrough
1220 }
1221 case Op_ShenandoahCompareAndSwapP:
1222 case Op_ShenandoahCompareAndSwapN:
1223 case Op_ShenandoahWeakCompareAndSwapP:
1224 case Op_ShenandoahWeakCompareAndSwapN:
1225 return conn_graph->add_final_edges_unsafe_access(n, opcode);
1226 case Op_ShenandoahEnqueueBarrier:
1227 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), NULL);
1228 return true;
1229 case Op_ShenandoahLoadReferenceBarrier:
1230 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), NULL);
1231 return true;
1232 default:
1233 // Nothing
1234 break;
1235 }
1236 return false;
1237 }
1238
1239 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1240 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1241 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1242
1243 }
1244
1245 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1246 switch (opcode) {
1247 case Op_ShenandoahCompareAndExchangeP:
1248 case Op_ShenandoahCompareAndExchangeN:
1249 case Op_ShenandoahWeakCompareAndSwapP:
1250 case Op_ShenandoahWeakCompareAndSwapN:
1251 case Op_ShenandoahCompareAndSwapP:
1252 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree
1253 Node* newval = n->in(MemNode::ValueIn);
1254 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1255 Node* pair = new BinaryNode(oldval, newval);
1256 n->set_req(MemNode::ValueIn,pair);
1257 n->del_req(LoadStoreConditionalNode::ExpectedIn);
1258 return true;
1259 }
1260 default:
1261 break;
1262 }
1263 return false;
1264 }
1265
1266 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1267 return xop == Op_ShenandoahCompareAndExchangeP ||
1268 xop == Op_ShenandoahCompareAndExchangeN ||
1269 xop == Op_ShenandoahWeakCompareAndSwapP ||
1270 xop == Op_ShenandoahWeakCompareAndSwapN ||
1271 xop == Op_ShenandoahCompareAndSwapN ||
1272 xop == Op_ShenandoahCompareAndSwapP;
1273 }