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 #include "opto/runtime.hpp"
42
43 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() {
44 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2());
45 }
46
47 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena)
48 : _enqueue_barriers(new (comp_arena) GrowableArray<ShenandoahEnqueueBarrierNode*>(comp_arena, 8, 0, NULL)),
49 _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8, 0, NULL)) {
50 }
51
52 int ShenandoahBarrierSetC2State::enqueue_barriers_count() const {
53 return _enqueue_barriers->length();
54 }
55
56 ShenandoahEnqueueBarrierNode* ShenandoahBarrierSetC2State::enqueue_barrier(int idx) const {
57 return _enqueue_barriers->at(idx);
58 }
59
60 void ShenandoahBarrierSetC2State::add_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) {
61 assert(!_enqueue_barriers->contains(n), "duplicate entry in barrier list");
62 _enqueue_barriers->append(n);
63 }
64
65 void ShenandoahBarrierSetC2State::remove_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) {
66 if (_enqueue_barriers->contains(n)) {
67 _enqueue_barriers->remove(n);
68 }
69 }
70
71 int ShenandoahBarrierSetC2State::load_reference_barriers_count() const {
72 return _load_reference_barriers->length();
73 }
74
75 ShenandoahLoadReferenceBarrierNode* ShenandoahBarrierSetC2State::load_reference_barrier(int idx) const {
76 return _load_reference_barriers->at(idx);
77 }
78
79 void ShenandoahBarrierSetC2State::add_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
80 assert(!_load_reference_barriers->contains(n), "duplicate entry in barrier list");
81 _load_reference_barriers->append(n);
82 }
83
84 void ShenandoahBarrierSetC2State::remove_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
85 if (_load_reference_barriers->contains(n)) {
86 _load_reference_barriers->remove(n);
87 }
88 }
89
90 Node* ShenandoahBarrierSetC2::shenandoah_storeval_barrier(GraphKit* kit, Node* obj) const {
91 if (ShenandoahStoreValEnqueueBarrier) {
92 obj = shenandoah_enqueue_barrier(kit, obj);
93 }
94 return obj;
95 }
96
97 #define __ kit->
98
99 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseTransform* phase, Node* adr,
100 BasicType bt, uint adr_idx) const {
101 intptr_t offset = 0;
102 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
103 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
104
105 if (offset == Type::OffsetBot) {
106 return false; // cannot unalias unless there are precise offsets
107 }
108
109 if (alloc == NULL) {
110 return false; // No allocation found
111 }
112
113 intptr_t size_in_bytes = type2aelembytes(bt);
114
115 Node* mem = __ memory(adr_idx); // start searching here...
116
117 for (int cnt = 0; cnt < 50; cnt++) {
118
119 if (mem->is_Store()) {
120
121 Node* st_adr = mem->in(MemNode::Address);
122 intptr_t st_offset = 0;
123 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
124
125 if (st_base == NULL) {
126 break; // inscrutable pointer
127 }
128
129 // Break we have found a store with same base and offset as ours so break
130 if (st_base == base && st_offset == offset) {
131 break;
132 }
133
134 if (st_offset != offset && st_offset != Type::OffsetBot) {
135 const int MAX_STORE = BytesPerLong;
136 if (st_offset >= offset + size_in_bytes ||
137 st_offset <= offset - MAX_STORE ||
138 st_offset <= offset - mem->as_Store()->memory_size()) {
139 // Success: The offsets are provably independent.
140 // (You may ask, why not just test st_offset != offset and be done?
141 // The answer is that stores of different sizes can co-exist
142 // in the same sequence of RawMem effects. We sometimes initialize
143 // a whole 'tile' of array elements with a single jint or jlong.)
144 mem = mem->in(MemNode::Memory);
145 continue; // advance through independent store memory
146 }
147 }
148
149 if (st_base != base
150 && MemNode::detect_ptr_independence(base, alloc, st_base,
151 AllocateNode::Ideal_allocation(st_base, phase),
152 phase)) {
153 // Success: The bases are provably independent.
154 mem = mem->in(MemNode::Memory);
155 continue; // advance through independent store memory
156 }
157 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
158
159 InitializeNode* st_init = mem->in(0)->as_Initialize();
160 AllocateNode* st_alloc = st_init->allocation();
161
162 // Make sure that we are looking at the same allocation site.
163 // The alloc variable is guaranteed to not be null here from earlier check.
164 if (alloc == st_alloc) {
165 // Check that the initialization is storing NULL so that no previous store
166 // has been moved up and directly write a reference
167 Node* captured_store = st_init->find_captured_store(offset,
168 type2aelembytes(T_OBJECT),
169 phase);
170 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
171 return true;
172 }
173 }
174 }
175
176 // Unless there is an explicit 'continue', we must bail out here,
177 // because 'mem' is an inscrutable memory state (e.g., a call).
178 break;
179 }
180
181 return false;
182 }
183
184 #undef __
185 #define __ ideal.
186
187 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit,
188 bool do_load,
189 Node* obj,
190 Node* adr,
191 uint alias_idx,
192 Node* val,
193 const TypeOopPtr* val_type,
194 Node* pre_val,
195 BasicType bt) const {
196 // Some sanity checks
197 // Note: val is unused in this routine.
198
199 if (do_load) {
200 // We need to generate the load of the previous value
201 assert(obj != NULL, "must have a base");
202 assert(adr != NULL, "where are loading from?");
203 assert(pre_val == NULL, "loaded already?");
204 assert(val_type != NULL, "need a type");
205
206 if (ReduceInitialCardMarks
207 && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
208 return;
209 }
210
211 } else {
212 // In this case both val_type and alias_idx are unused.
213 assert(pre_val != NULL, "must be loaded already");
214 // Nothing to be done if pre_val is null.
215 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
216 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
217 }
218 assert(bt == T_OBJECT, "or we shouldn't be here");
219
220 IdealKit ideal(kit, true);
221
222 Node* tls = __ thread(); // ThreadLocalStorage
223
224 Node* no_base = __ top();
225 Node* zero = __ ConI(0);
226 Node* zeroX = __ ConX(0);
227
228 float likely = PROB_LIKELY(0.999);
229 float unlikely = PROB_UNLIKELY(0.999);
230
231 // Offsets into the thread
232 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
233 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
234
235 // Now the actual pointers into the thread
236 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
237 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
238
239 // Now some of the values
240 Node* marking;
241 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())));
242 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
243 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
244 assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape");
245
246 // if (!marking)
247 __ if_then(marking, BoolTest::ne, zero, unlikely); {
248 BasicType index_bt = TypeX_X->basic_type();
249 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
250 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
251
252 if (do_load) {
253 // load original value
254 // alias_idx correct??
255 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
256 }
257
258 // if (pre_val != NULL)
259 __ if_then(pre_val, BoolTest::ne, kit->null()); {
260 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
261
262 // is the queue for this thread full?
263 __ if_then(index, BoolTest::ne, zeroX, likely); {
264
265 // decrement the index
266 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
267
268 // Now get the buffer location we will log the previous value into and store it
269 Node *log_addr = __ AddP(no_base, buffer, next_index);
270 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
271 // update the index
272 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
273
274 } __ else_(); {
275
276 // logging buffer is full, call the runtime
277 const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type();
278 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls);
279 } __ end_if(); // (!index)
280 } __ end_if(); // (pre_val != NULL)
281 } __ end_if(); // (!marking)
282
283 // Final sync IdealKit and GraphKit.
284 kit->final_sync(ideal);
285
286 if (ShenandoahSATBBarrier && adr != NULL) {
287 Node* c = kit->control();
288 Node* call = c->in(1)->in(1)->in(1)->in(0);
289 assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected");
290 call->add_req(adr);
291 }
292 }
293
294 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) {
295 return call->is_CallLeaf() &&
296 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry);
297 }
298
299 bool ShenandoahBarrierSetC2::is_shenandoah_lrb_call(Node* call) {
300 if (!call->is_CallLeaf()) {
301 return false;
302 }
303
304 address entry_point = call->as_CallLeaf()->entry_point();
305 return (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier)) ||
306 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_narrow)) ||
307 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_native));
308 }
309
310 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseTransform *phase, Node* n) {
311 if (n->Opcode() != Op_If) {
312 return false;
313 }
314
315 Node* bol = n->in(1);
316 assert(bol->is_Bool(), "");
317 Node* cmpx = bol->in(1);
318 if (bol->as_Bool()->_test._test == BoolTest::ne &&
319 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
320 is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
321 cmpx->in(1)->in(2)->is_Con() &&
322 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
323 return true;
324 }
325
326 return false;
327 }
328
329 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
330 if (!n->is_Load()) return false;
331 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
332 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
333 && n->in(2)->in(3)->is_Con()
334 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
335 }
336
337 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
338 bool do_load,
339 Node* obj,
340 Node* adr,
341 uint alias_idx,
342 Node* val,
343 const TypeOopPtr* val_type,
344 Node* pre_val,
345 BasicType bt) const {
346 if (ShenandoahSATBBarrier) {
347 IdealKit ideal(kit);
348 kit->sync_kit(ideal);
349
350 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
351
352 ideal.sync_kit(kit);
353 kit->final_sync(ideal);
354 }
355 }
356
357 Node* ShenandoahBarrierSetC2::shenandoah_enqueue_barrier(GraphKit* kit, Node* pre_val) const {
358 return kit->gvn().transform(new ShenandoahEnqueueBarrierNode(pre_val));
359 }
360
361 // Helper that guards and inserts a pre-barrier.
362 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
363 Node* pre_val, bool need_mem_bar) const {
364 // We could be accessing the referent field of a reference object. If so, when G1
365 // is enabled, we need to log the value in the referent field in an SATB buffer.
366 // This routine performs some compile time filters and generates suitable
367 // runtime filters that guard the pre-barrier code.
368 // Also add memory barrier for non volatile load from the referent field
369 // to prevent commoning of loads across safepoint.
370
371 // Some compile time checks.
372
373 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
374 const TypeX* otype = offset->find_intptr_t_type();
375 if (otype != NULL && otype->is_con() &&
376 otype->get_con() != java_lang_ref_Reference::referent_offset) {
377 // Constant offset but not the reference_offset so just return
378 return;
379 }
380
381 // We only need to generate the runtime guards for instances.
382 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
383 if (btype != NULL) {
384 if (btype->isa_aryptr()) {
385 // Array type so nothing to do
386 return;
387 }
388
389 const TypeInstPtr* itype = btype->isa_instptr();
390 if (itype != NULL) {
391 // Can the klass of base_oop be statically determined to be
392 // _not_ a sub-class of Reference and _not_ Object?
393 ciKlass* klass = itype->klass();
394 if ( klass->is_loaded() &&
395 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
396 !kit->env()->Object_klass()->is_subtype_of(klass)) {
397 return;
398 }
399 }
400 }
401
402 // The compile time filters did not reject base_oop/offset so
403 // we need to generate the following runtime filters
404 //
405 // if (offset == java_lang_ref_Reference::_reference_offset) {
406 // if (instance_of(base, java.lang.ref.Reference)) {
407 // pre_barrier(_, pre_val, ...);
408 // }
409 // }
410
411 float likely = PROB_LIKELY( 0.999);
412 float unlikely = PROB_UNLIKELY(0.999);
413
414 IdealKit ideal(kit);
415
416 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
417
418 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
419 // Update graphKit memory and control from IdealKit.
420 kit->sync_kit(ideal);
421
422 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
423 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
424
425 // Update IdealKit memory and control from graphKit.
426 __ sync_kit(kit);
427
428 Node* one = __ ConI(1);
429 // is_instof == 0 if base_oop == NULL
430 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
431
432 // Update graphKit from IdeakKit.
433 kit->sync_kit(ideal);
434
435 // Use the pre-barrier to record the value in the referent field
436 satb_write_barrier_pre(kit, false /* do_load */,
437 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
438 pre_val /* pre_val */,
439 T_OBJECT);
440 if (need_mem_bar) {
441 // Add memory barrier to prevent commoning reads from this field
442 // across safepoint since GC can change its value.
443 kit->insert_mem_bar(Op_MemBarCPUOrder);
444 }
445 // Update IdealKit from graphKit.
446 __ sync_kit(kit);
447
448 } __ end_if(); // _ref_type != ref_none
449 } __ end_if(); // offset == referent_offset
450
451 // Final sync IdealKit and GraphKit.
452 kit->final_sync(ideal);
453 }
454
455 #undef __
456
457 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() {
458 const Type **fields = TypeTuple::fields(2);
459 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
460 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
461 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
462
463 // create result type (range)
464 fields = TypeTuple::fields(0);
465 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
466
467 return TypeFunc::make(domain, range);
468 }
469
470 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() {
471 const Type **fields = TypeTuple::fields(1);
472 fields[TypeFunc::Parms+0] = TypeOopPtr::NOTNULL; // src oop
473 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
474
475 // create result type (range)
476 fields = TypeTuple::fields(0);
477 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
478
479 return TypeFunc::make(domain, range);
480 }
481
482 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type() {
483 const Type **fields = TypeTuple::fields(2);
484 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
485 fields[TypeFunc::Parms+1] = TypeRawPtr::BOTTOM; // original load address
486
487 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
488
489 // create result type (range)
490 fields = TypeTuple::fields(1);
491 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL;
492 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
493
494 return TypeFunc::make(domain, range);
495 }
496
497 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
498 DecoratorSet decorators = access.decorators();
499
500 const TypePtr* adr_type = access.addr().type();
501 Node* adr = access.addr().node();
502
503 bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
504 bool on_heap = (decorators & IN_HEAP) != 0;
505
506 if (!access.is_oop() || (!on_heap && !anonymous)) {
507 return BarrierSetC2::store_at_resolved(access, val);
508 }
509
510 if (access.is_parse_access()) {
511 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
512 GraphKit* kit = parse_access.kit();
513
514 uint adr_idx = kit->C->get_alias_index(adr_type);
515 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
516 Node* value = val.node();
517 value = shenandoah_storeval_barrier(kit, value);
518 val.set_node(value);
519 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
520 static_cast<const TypeOopPtr*>(val.type()), NULL /* pre_val */, access.type());
521 } else {
522 assert(access.is_opt_access(), "only for optimization passes");
523 assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code");
524 C2OptAccess& opt_access = static_cast<C2OptAccess&>(access);
525 PhaseGVN& gvn = opt_access.gvn();
526 MergeMemNode* mm = opt_access.mem();
527
528 if (ShenandoahStoreValEnqueueBarrier) {
529 Node* enqueue = gvn.transform(new ShenandoahEnqueueBarrierNode(val.node()));
530 val.set_node(enqueue);
531 }
532 }
533 return BarrierSetC2::store_at_resolved(access, val);
534 }
535
536 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
537 DecoratorSet decorators = access.decorators();
538 BasicType type = access.type();
539
540 assert((decorators & AS_RAW) == 0, "Unexpected decorator");
541 assert((decorators & AS_NO_KEEPALIVE) == 0, "Unexpected decorator");
542
543 Node* adr = access.addr().node();
544 Node* obj = access.base();
545
546 bool mismatched = (decorators & C2_MISMATCHED) != 0;
547 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
548 bool on_heap = (decorators & IN_HEAP) != 0;
549 bool on_weak_ref = (decorators & (ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF)) != 0;
550 bool is_unordered = (decorators & MO_UNORDERED) != 0;
551 bool need_cpu_mem_bar = !is_unordered || mismatched || !on_heap;
552
553 Node* top = Compile::current()->top();
554
555 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
556 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
557 if (!ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
558 return load;
559 }
560
561 load = new ShenandoahLoadReferenceBarrierNode(NULL, load, ShenandoahBarrierSet::use_native_load_reference_barrier(decorators, type));
562 if (access.is_parse_access()) {
563 load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load);
564 } else {
565 load = static_cast<C2OptAccess &>(access).gvn().transform(load);
566 }
567
568 if (!ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
569 return load;
570 }
571 // If we are reading the value of the referent field of a Reference
572 // object (either by using Unsafe directly or through reflection)
573 // then, if SATB is enabled, we need to record the referent in an
574 // SATB log buffer using the pre-barrier mechanism.
575 // Also we need to add memory barrier to prevent commoning reads
576 // from this field across safepoint since GC can change its value.
577 if (unknown && (offset == top || obj == top)) {
578 return load;
579 }
580
581 assert(access.is_parse_access(), "entry not supported at optimization time");
582 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
583 GraphKit* kit = parse_access.kit();
584
585 if (on_weak_ref) {
586 // Use the pre-barrier to record the value in the referent field
587 satb_write_barrier_pre(kit, false /* do_load */,
588 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
589 load /* pre_val */, T_OBJECT);
590 // Add memory barrier to prevent commoning reads from this field
591 // across safepoint since GC can change its value.
592 kit->insert_mem_bar(Op_MemBarCPUOrder);
593 } else if (unknown) {
594 // We do not require a mem bar inside pre_barrier if need_mem_bar
595 // is set: the barriers would be emitted by us.
596 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
597 }
598
599 return load;
600 }
601
602 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
603 Node* new_val, const Type* value_type) const {
604 GraphKit* kit = access.kit();
605 if (access.is_oop()) {
606 new_val = shenandoah_storeval_barrier(kit, new_val);
607 shenandoah_write_barrier_pre(kit, false /* do_load */,
608 NULL, NULL, max_juint, NULL, NULL,
609 expected_val /* pre_val */, T_OBJECT);
610
611 MemNode::MemOrd mo = access.mem_node_mo();
612 Node* mem = access.memory();
613 Node* adr = access.addr().node();
614 const TypePtr* adr_type = access.addr().type();
615 Node* load_store = NULL;
616
617 #ifdef _LP64
618 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
619 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
620 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
621 if (ShenandoahCASBarrier) {
622 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
623 } else {
624 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
625 }
626 } else
627 #endif
628 {
629 if (ShenandoahCASBarrier) {
630 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
631 } else {
632 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
633 }
634 }
635
636 access.set_raw_access(load_store);
637 pin_atomic_op(access);
638
639 #ifdef _LP64
640 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
641 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
642 }
643 #endif
644 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, load_store, false));
645 return load_store;
646 }
647 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
648 }
649
650 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
651 Node* new_val, const Type* value_type) const {
652 GraphKit* kit = access.kit();
653 if (access.is_oop()) {
654 new_val = shenandoah_storeval_barrier(kit, new_val);
655 shenandoah_write_barrier_pre(kit, false /* do_load */,
656 NULL, NULL, max_juint, NULL, NULL,
657 expected_val /* pre_val */, T_OBJECT);
658 DecoratorSet decorators = access.decorators();
659 MemNode::MemOrd mo = access.mem_node_mo();
660 Node* mem = access.memory();
661 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
662 Node* load_store = NULL;
663 Node* adr = access.addr().node();
664 #ifdef _LP64
665 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
666 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
667 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
668 if (ShenandoahCASBarrier) {
669 if (is_weak_cas) {
670 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
671 } else {
672 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
673 }
674 } else {
675 if (is_weak_cas) {
676 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
677 } else {
678 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
679 }
680 }
681 } else
682 #endif
683 {
684 if (ShenandoahCASBarrier) {
685 if (is_weak_cas) {
686 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
687 } else {
688 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
689 }
690 } else {
691 if (is_weak_cas) {
692 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
693 } else {
694 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
695 }
696 }
697 }
698 access.set_raw_access(load_store);
699 pin_atomic_op(access);
700 return load_store;
701 }
702 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
703 }
704
705 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
706 GraphKit* kit = access.kit();
707 if (access.is_oop()) {
708 val = shenandoah_storeval_barrier(kit, val);
709 }
710 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
711 if (access.is_oop()) {
712 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, result, false));
713 shenandoah_write_barrier_pre(kit, false /* do_load */,
714 NULL, NULL, max_juint, NULL, NULL,
715 result /* pre_val */, T_OBJECT);
716 }
717 return result;
718 }
719
720 // Support for GC barriers emitted during parsing
721 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
722 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) return true;
723 if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) {
724 return false;
725 }
726 CallLeafNode *call = node->as_CallLeaf();
727 if (call->_name == NULL) {
728 return false;
729 }
730
731 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 ||
732 strcmp(call->_name, "shenandoah_cas_obj") == 0 ||
733 strcmp(call->_name, "shenandoah_wb_pre") == 0;
734 }
735
736 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
737 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
738 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
739 }
740 if (c->Opcode() == Op_ShenandoahEnqueueBarrier) {
741 c = c->in(1);
742 }
743 return c;
744 }
745
746 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
747 return !ShenandoahBarrierC2Support::expand(C, igvn);
748 }
749
750 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
751 if (mode == LoopOptsShenandoahExpand) {
752 assert(UseShenandoahGC, "only for shenandoah");
753 ShenandoahBarrierC2Support::pin_and_expand(phase);
754 return true;
755 } else if (mode == LoopOptsShenandoahPostExpand) {
756 assert(UseShenandoahGC, "only for shenandoah");
757 visited.Clear();
758 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
759 return true;
760 }
761 return false;
762 }
763
764 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, ArrayCopyPhase phase) const {
765 bool is_oop = is_reference_type(type);
766 if (!is_oop) {
767 return false;
768 }
769 if (tightly_coupled_alloc) {
770 if (phase == Optimization) {
771 return false;
772 }
773 return !is_clone;
774 }
775 if (phase == Optimization) {
776 return !ShenandoahStoreValEnqueueBarrier;
777 }
778 return true;
779 }
780
781 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
782 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
783 if (src_type->isa_instptr() != NULL) {
784 ciInstanceKlass* ik = src_type->klass()->as_instance_klass();
785 if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) {
786 if (ik->has_object_fields()) {
787 return true;
788 } else {
789 if (!src_type->klass_is_exact()) {
790 Compile::current()->dependencies()->assert_leaf_type(ik);
791 }
792 }
793 } else {
794 return true;
795 }
796 } else if (src_type->isa_aryptr()) {
797 BasicType src_elem = src_type->klass()->as_array_klass()->element_type()->basic_type();
798 if (is_reference_type(src_elem)) {
799 return true;
800 }
801 } else {
802 return true;
803 }
804 return false;
805 }
806
807 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
808 Node* ctrl = ac->in(TypeFunc::Control);
809 Node* mem = ac->in(TypeFunc::Memory);
810 Node* src = ac->in(ArrayCopyNode::Src);
811 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
812 Node* dest = ac->in(ArrayCopyNode::Dest);
813 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
814 Node* length = ac->in(ArrayCopyNode::Length);
815 assert (src_offset == NULL && dest_offset == NULL, "for clone offsets should be null");
816 assert (src->is_AddP(), "for clone the src should be the interior ptr");
817 assert (dest->is_AddP(), "for clone the dst should be the interior ptr");
818
819 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
820 // Check if heap is has forwarded objects. If it does, we need to call into the special
821 // routine that would fix up source references before we can continue.
822
823 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
824 Node* region = new RegionNode(PATH_LIMIT);
825 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM);
826
827 Node* thread = phase->transform_later(new ThreadLocalNode());
828 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
829 Node* gc_state_addr = phase->transform_later(new AddPNode(phase->C->top(), thread, offset));
830
831 uint gc_state_idx = Compile::AliasIdxRaw;
832 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument
833 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
834
835 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered));
836 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED)));
837 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT)));
838 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne));
839
840 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If();
841 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff));
842 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff));
843
844 // Heap is stable, no need to do anything additional
845 region->init_req(_heap_stable, stable_ctrl);
846 mem_phi->init_req(_heap_stable, mem);
847
848 // Heap is unstable, call into clone barrier stub
849 Node* call = phase->make_leaf_call(unstable_ctrl, mem,
850 ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(),
851 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier),
852 "shenandoah_clone",
853 TypeRawPtr::BOTTOM,
854 src->in(AddPNode::Base));
855 call = phase->transform_later(call);
856
857 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control));
858 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory));
859 region->init_req(_heap_unstable, ctrl);
860 mem_phi->init_req(_heap_unstable, mem);
861
862 // Wire up the actual arraycopy stub now
863 ctrl = phase->transform_later(region);
864 mem = phase->transform_later(mem_phi);
865
866 const char* name = "arraycopy";
867 call = phase->make_leaf_call(ctrl, mem,
868 OptoRuntime::fast_arraycopy_Type(),
869 phase->basictype2arraycopy(T_LONG, NULL, NULL, true, name, true),
870 name, TypeRawPtr::BOTTOM,
871 src, dest, length
872 LP64_ONLY(COMMA phase->top()));
873 call = phase->transform_later(call);
874
875 // Hook up the whole thing into the graph
876 phase->igvn().replace_node(ac, call);
877 } else {
878 BarrierSetC2::clone_at_expansion(phase, ac);
879 }
880 }
881
882
883 // Support for macro expanded GC barriers
884 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
885 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
886 state()->add_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
887 }
888 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
889 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
890 }
891 }
892
893 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
894 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
895 state()->remove_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
896 }
897 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
898 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
899 }
900 }
901
902 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const {
903 if (is_shenandoah_wb_pre_call(n)) {
904 shenandoah_eliminate_wb_pre(n, ¯o->igvn());
905 }
906 }
907
908 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
909 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
910 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
911 c = c->unique_ctrl_out();
912 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
913 c = c->unique_ctrl_out();
914 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
915 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
916 assert(iff->is_If(), "expect test");
917 if (!is_shenandoah_marking_if(igvn, iff)) {
918 c = c->unique_ctrl_out();
919 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
920 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
921 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
922 }
923 Node* cmpx = iff->in(1)->in(1);
924 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
925 igvn->rehash_node_delayed(call);
926 call->del_req(call->req()-1);
927 }
928
929 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
930 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
931 igvn->add_users_to_worklist(node);
932 }
933 }
934
935 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
936 for (uint i = 0; i < useful.size(); i++) {
937 Node* n = useful.at(i);
938 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
939 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
940 C->record_for_igvn(n->fast_out(i));
941 }
942 }
943 }
944 for (int i = state()->enqueue_barriers_count() - 1; i >= 0; i--) {
945 ShenandoahEnqueueBarrierNode* n = state()->enqueue_barrier(i);
946 if (!useful.member(n)) {
947 state()->remove_enqueue_barrier(n);
948 }
949 }
950 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
951 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
952 if (!useful.member(n)) {
953 state()->remove_load_reference_barrier(n);
954 }
955 }
956 }
957
958 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
959 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
960 }
961
962 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
963 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
964 }
965
966 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
967 // expanded later, then now is the time to do so.
968 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
969
970 #ifdef ASSERT
971 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
972 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) {
973 ShenandoahBarrierC2Support::verify(Compile::current()->root());
974 } else if (phase == BarrierSetC2::BeforeCodeGen) {
975 // Verify G1 pre-barriers
976 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset());
977
978 ResourceArea *area = Thread::current()->resource_area();
979 Unique_Node_List visited(area);
980 Node_List worklist(area);
981 // We're going to walk control flow backwards starting from the Root
982 worklist.push(compile->root());
983 while (worklist.size() > 0) {
984 Node *x = worklist.pop();
985 if (x == NULL || x == compile->top()) continue;
986 if (visited.member(x)) {
987 continue;
988 } else {
989 visited.push(x);
990 }
991
992 if (x->is_Region()) {
993 for (uint i = 1; i < x->req(); i++) {
994 worklist.push(x->in(i));
995 }
996 } else {
997 worklist.push(x->in(0));
998 // We are looking for the pattern:
999 // /->ThreadLocal
1000 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
1001 // \->ConI(0)
1002 // We want to verify that the If and the LoadB have the same control
1003 // See GraphKit::g1_write_barrier_pre()
1004 if (x->is_If()) {
1005 IfNode *iff = x->as_If();
1006 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
1007 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
1008 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
1009 && cmp->in(1)->is_Load()) {
1010 LoadNode *load = cmp->in(1)->as_Load();
1011 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
1012 && load->in(2)->in(3)->is_Con()
1013 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
1014
1015 Node *if_ctrl = iff->in(0);
1016 Node *load_ctrl = load->in(0);
1017
1018 if (if_ctrl != load_ctrl) {
1019 // Skip possible CProj->NeverBranch in infinite loops
1020 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
1021 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
1022 if_ctrl = if_ctrl->in(0)->in(0);
1023 }
1024 }
1025 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
1026 }
1027 }
1028 }
1029 }
1030 }
1031 }
1032 }
1033 }
1034 #endif
1035
1036 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
1037 if (is_shenandoah_wb_pre_call(n)) {
1038 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1039 if (n->req() > cnt) {
1040 Node* addp = n->in(cnt);
1041 if (has_only_shenandoah_wb_pre_uses(addp)) {
1042 n->del_req(cnt);
1043 if (can_reshape) {
1044 phase->is_IterGVN()->_worklist.push(addp);
1045 }
1046 return n;
1047 }
1048 }
1049 }
1050 if (n->Opcode() == Op_CmpP) {
1051 Node* in1 = n->in(1);
1052 Node* in2 = n->in(2);
1053 if (in1->bottom_type() == TypePtr::NULL_PTR) {
1054 in2 = step_over_gc_barrier(in2);
1055 }
1056 if (in2->bottom_type() == TypePtr::NULL_PTR) {
1057 in1 = step_over_gc_barrier(in1);
1058 }
1059 PhaseIterGVN* igvn = phase->is_IterGVN();
1060 if (in1 != n->in(1)) {
1061 if (igvn != NULL) {
1062 n->set_req_X(1, in1, igvn);
1063 } else {
1064 n->set_req(1, in1);
1065 }
1066 assert(in2 == n->in(2), "only one change");
1067 return n;
1068 }
1069 if (in2 != n->in(2)) {
1070 if (igvn != NULL) {
1071 n->set_req_X(2, in2, igvn);
1072 } else {
1073 n->set_req(2, in2);
1074 }
1075 return n;
1076 }
1077 } else if (can_reshape &&
1078 n->Opcode() == Op_If &&
1079 ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1080 n->in(0) != NULL) {
1081 Node* dom = n->in(0);
1082 Node* prev_dom = n;
1083 int op = n->Opcode();
1084 int dist = 16;
1085 // Search up the dominator tree for another heap stable test
1086 while (dom->Opcode() != op || // Not same opcode?
1087 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1?
1088 prev_dom->in(0) != dom) { // One path of test does not dominate?
1089 if (dist < 0) return NULL;
1090
1091 dist--;
1092 prev_dom = dom;
1093 dom = IfNode::up_one_dom(dom);
1094 if (!dom) return NULL;
1095 }
1096
1097 // Check that we did not follow a loop back to ourselves
1098 if (n == dom) {
1099 return NULL;
1100 }
1101
1102 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN());
1103 }
1104
1105 return NULL;
1106 }
1107
1108 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1109 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1110 Node* u = n->fast_out(i);
1111 if (!is_shenandoah_wb_pre_call(u)) {
1112 return false;
1113 }
1114 }
1115 return n->outcnt() > 0;
1116 }
1117
1118 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode) const {
1119 switch (opcode) {
1120 case Op_CallLeaf:
1121 case Op_CallLeafNoFP: {
1122 assert (n->is_Call(), "");
1123 CallNode *call = n->as_Call();
1124 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1125 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1126 if (call->req() > cnt) {
1127 assert(call->req() == cnt + 1, "only one extra input");
1128 Node *addp = call->in(cnt);
1129 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1130 call->del_req(cnt);
1131 }
1132 }
1133 return false;
1134 }
1135 case Op_ShenandoahCompareAndSwapP:
1136 case Op_ShenandoahCompareAndSwapN:
1137 case Op_ShenandoahWeakCompareAndSwapN:
1138 case Op_ShenandoahWeakCompareAndSwapP:
1139 case Op_ShenandoahCompareAndExchangeP:
1140 case Op_ShenandoahCompareAndExchangeN:
1141 #ifdef ASSERT
1142 if( VerifyOptoOopOffsets ) {
1143 MemNode* mem = n->as_Mem();
1144 // Check to see if address types have grounded out somehow.
1145 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
1146 ciInstanceKlass *k = tp->klass()->as_instance_klass();
1147 bool oop_offset_is_sane = k->contains_field_offset(tp->offset());
1148 assert( !tp || oop_offset_is_sane, "" );
1149 }
1150 #endif
1151 return true;
1152 case Op_ShenandoahLoadReferenceBarrier:
1153 assert(false, "should have been expanded already");
1154 return true;
1155 default:
1156 return false;
1157 }
1158 }
1159
1160 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1161 switch (opcode) {
1162 case Op_ShenandoahCompareAndExchangeP:
1163 case Op_ShenandoahCompareAndExchangeN:
1164 conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1165 // fallthrough
1166 case Op_ShenandoahWeakCompareAndSwapP:
1167 case Op_ShenandoahWeakCompareAndSwapN:
1168 case Op_ShenandoahCompareAndSwapP:
1169 case Op_ShenandoahCompareAndSwapN:
1170 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1171 return true;
1172 case Op_StoreP: {
1173 Node* adr = n->in(MemNode::Address);
1174 const Type* adr_type = gvn->type(adr);
1175 // Pointer stores in G1 barriers looks like unsafe access.
1176 // Ignore such stores to be able scalar replace non-escaping
1177 // allocations.
1178 if (adr_type->isa_rawptr() && adr->is_AddP()) {
1179 Node* base = conn_graph->get_addp_base(adr);
1180 if (base->Opcode() == Op_LoadP &&
1181 base->in(MemNode::Address)->is_AddP()) {
1182 adr = base->in(MemNode::Address);
1183 Node* tls = conn_graph->get_addp_base(adr);
1184 if (tls->Opcode() == Op_ThreadLocal) {
1185 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1186 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1187 if (offs == buf_offset) {
1188 return true; // Pre barrier previous oop value store.
1189 }
1190 }
1191 }
1192 }
1193 return false;
1194 }
1195 case Op_ShenandoahEnqueueBarrier:
1196 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1197 break;
1198 case Op_ShenandoahLoadReferenceBarrier:
1199 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist);
1200 return true;
1201 default:
1202 // Nothing
1203 break;
1204 }
1205 return false;
1206 }
1207
1208 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1209 switch (opcode) {
1210 case Op_ShenandoahCompareAndExchangeP:
1211 case Op_ShenandoahCompareAndExchangeN: {
1212 Node *adr = n->in(MemNode::Address);
1213 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
1214 // fallthrough
1215 }
1216 case Op_ShenandoahCompareAndSwapP:
1217 case Op_ShenandoahCompareAndSwapN:
1218 case Op_ShenandoahWeakCompareAndSwapP:
1219 case Op_ShenandoahWeakCompareAndSwapN:
1220 return conn_graph->add_final_edges_unsafe_access(n, opcode);
1221 case Op_ShenandoahEnqueueBarrier:
1222 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), NULL);
1223 return true;
1224 case Op_ShenandoahLoadReferenceBarrier:
1225 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), NULL);
1226 return true;
1227 default:
1228 // Nothing
1229 break;
1230 }
1231 return false;
1232 }
1233
1234 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1235 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1236 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1237
1238 }
1239
1240 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1241 switch (opcode) {
1242 case Op_ShenandoahCompareAndExchangeP:
1243 case Op_ShenandoahCompareAndExchangeN:
1244 case Op_ShenandoahWeakCompareAndSwapP:
1245 case Op_ShenandoahWeakCompareAndSwapN:
1246 case Op_ShenandoahCompareAndSwapP:
1247 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree
1248 Node* newval = n->in(MemNode::ValueIn);
1249 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1250 Node* pair = new BinaryNode(oldval, newval);
1251 n->set_req(MemNode::ValueIn,pair);
1252 n->del_req(LoadStoreConditionalNode::ExpectedIn);
1253 return true;
1254 }
1255 default:
1256 break;
1257 }
1258 return false;
1259 }
1260
1261 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1262 return xop == Op_ShenandoahCompareAndExchangeP ||
1263 xop == Op_ShenandoahCompareAndExchangeN ||
1264 xop == Op_ShenandoahWeakCompareAndSwapP ||
1265 xop == Op_ShenandoahWeakCompareAndSwapN ||
1266 xop == Op_ShenandoahCompareAndSwapN ||
1267 xop == Op_ShenandoahCompareAndSwapP;
1268 }