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/shenandoahBarrierSet.hpp"
27 #include "gc/shenandoah/shenandoahForwarding.hpp"
28 #include "gc/shenandoah/shenandoahHeap.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 "gc/shenandoah/heuristics/shenandoahHeuristics.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 Shenandoah SATBMarkQueue::_index with wrong size.");
251 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
252
253 if (do_load) {
254 // load original value
255 // alias_idx correct??
256 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
257 }
258
259 // if (pre_val != NULL)
260 __ if_then(pre_val, BoolTest::ne, kit->null()); {
261 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
262
263 // is the queue for this thread full?
264 __ if_then(index, BoolTest::ne, zeroX, likely); {
265
266 // decrement the index
267 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
268
269 // Now get the buffer location we will log the previous value into and store it
270 Node *log_addr = __ AddP(no_base, buffer, next_index);
271 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
272 // update the index
273 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
274
275 } __ else_(); {
276
277 // logging buffer is full, call the runtime
278 const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type();
279 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls);
280 } __ end_if(); // (!index)
281 } __ end_if(); // (pre_val != NULL)
282 } __ end_if(); // (!marking)
283
284 // Final sync IdealKit and GraphKit.
285 kit->final_sync(ideal);
286
287 if (ShenandoahSATBBarrier && adr != NULL) {
288 Node* c = kit->control();
289 Node* call = c->in(1)->in(1)->in(1)->in(0);
290 assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected");
291 call->add_req(adr);
292 }
293 }
294
295 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) {
296 return call->is_CallLeaf() &&
297 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry);
298 }
299
300 bool ShenandoahBarrierSetC2::is_shenandoah_lrb_call(Node* call) {
301 if (!call->is_CallLeaf()) {
302 return false;
303 }
304
305 address entry_point = call->as_CallLeaf()->entry_point();
306 return (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier)) ||
307 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_narrow));
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 GraphKit* kit = access.kit();
511
512 uint adr_idx = kit->C->get_alias_index(adr_type);
513 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
514 Node* value = val.node();
515 value = shenandoah_storeval_barrier(kit, value);
516 val.set_node(value);
517 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
518 static_cast<const TypeOopPtr*>(val.type()), NULL /* pre_val */, access.type());
519 return BarrierSetC2::store_at_resolved(access, val);
520 }
521
522 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
523 // 1: non-reference load, no additional barrier is needed
524 if (!access.is_oop()) {
525 return BarrierSetC2::load_at_resolved(access, val_type);;
526 }
527
528 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
529 DecoratorSet decorators = access.decorators();
530 BasicType type = access.type();
531
532 // 2: apply LRB if needed
533 if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
534 load = new ShenandoahLoadReferenceBarrierNode(NULL, load);
535 load = access.kit()->gvn().transform(load);
536 }
537
538 // 3: apply keep-alive barrier if needed
539 if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
540 Node* top = Compile::current()->top();
541 Node* adr = access.addr().node();
542 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
543 Node* obj = access.base();
544
545 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
546 bool on_weak_ref = (decorators & (ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF)) != 0;
547 bool keep_alive = (decorators & AS_NO_KEEPALIVE) == 0;
548
549 // If we are reading the value of the referent field of a Reference
550 // object (either by using Unsafe directly or through reflection)
551 // then, if SATB is enabled, we need to record the referent in an
552 // SATB log buffer using the pre-barrier mechanism.
553 // Also we need to add memory barrier to prevent commoning reads
554 // from this field across safepoint since GC can change its value.
555 if (!on_weak_ref || (unknown && (offset == top || obj == top)) || !keep_alive) {
556 return load;
557 }
558 GraphKit* kit = access.kit();
559 bool mismatched = (decorators & C2_MISMATCHED) != 0;
560 bool is_unordered = (decorators & MO_UNORDERED) != 0;
561 bool need_cpu_mem_bar = !is_unordered || mismatched;
562
563 if (on_weak_ref) {
564 // Use the pre-barrier to record the value in the referent field
565 satb_write_barrier_pre(kit, false /* do_load */,
566 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
567 load /* pre_val */, T_OBJECT);
568 // Add memory barrier to prevent commoning reads from this field
569 // across safepoint since GC can change its value.
570 kit->insert_mem_bar(Op_MemBarCPUOrder);
571 } else if (unknown) {
572 // We do not require a mem bar inside pre_barrier if need_mem_bar
573 // is set: the barriers would be emitted by us.
574 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
575 }
576 }
577
578 return load;
579 }
580
581 static void pin_atomic_op(C2AtomicAccess& access) {
582 if (!access.needs_pinning()) {
583 return;
584 }
585 // SCMemProjNodes represent the memory state of a LoadStore. Their
586 // main role is to prevent LoadStore nodes from being optimized away
587 // when their results aren't used.
588 GraphKit* kit = access.kit();
589 Node* load_store = access.raw_access();
590 assert(load_store != NULL, "must pin atomic op");
591 Node* proj = kit->gvn().transform(new SCMemProjNode(load_store));
592 kit->set_memory(proj, access.alias_idx());
593 }
594
595 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicAccess& access, Node* expected_val,
596 Node* new_val, const Type* value_type) const {
597 GraphKit* kit = access.kit();
598 if (access.is_oop()) {
599 new_val = shenandoah_storeval_barrier(kit, new_val);
600 shenandoah_write_barrier_pre(kit, false /* do_load */,
601 NULL, NULL, max_juint, NULL, NULL,
602 expected_val /* pre_val */, T_OBJECT);
603
604 MemNode::MemOrd mo = access.mem_node_mo();
605 Node* mem = access.memory();
606 Node* adr = access.addr().node();
607 const TypePtr* adr_type = access.addr().type();
608 Node* load_store = NULL;
609
610 #ifdef _LP64
611 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
612 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
613 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
614 if (ShenandoahCASBarrier) {
615 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
616 } else {
617 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
618 }
619 } else
620 #endif
621 {
622 if (ShenandoahCASBarrier) {
623 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
624 } else {
625 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
626 }
627 }
628
629 access.set_raw_access(load_store);
630 pin_atomic_op(access);
631
632 #ifdef _LP64
633 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
634 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
635 }
636 #endif
637 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, load_store));
638 return load_store;
639 }
640 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
641 }
642
643 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicAccess& access, Node* expected_val,
644 Node* new_val, const Type* value_type) const {
645 GraphKit* kit = access.kit();
646 if (access.is_oop()) {
647 new_val = shenandoah_storeval_barrier(kit, new_val);
648 shenandoah_write_barrier_pre(kit, false /* do_load */,
649 NULL, NULL, max_juint, NULL, NULL,
650 expected_val /* pre_val */, T_OBJECT);
651 DecoratorSet decorators = access.decorators();
652 MemNode::MemOrd mo = access.mem_node_mo();
653 Node* mem = access.memory();
654 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
655 Node* load_store = NULL;
656 Node* adr = access.addr().node();
657 #ifdef _LP64
658 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
659 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
660 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
661 if (ShenandoahCASBarrier) {
662 if (is_weak_cas) {
663 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
664 } else {
665 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
666 }
667 } else {
668 if (is_weak_cas) {
669 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
670 } else {
671 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
672 }
673 }
674 } else
675 #endif
676 {
677 if (ShenandoahCASBarrier) {
678 if (is_weak_cas) {
679 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
680 } else {
681 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
682 }
683 } else {
684 if (is_weak_cas) {
685 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
686 } else {
687 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
688 }
689 }
690 }
691 access.set_raw_access(load_store);
692 pin_atomic_op(access);
693 return load_store;
694 }
695 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
696 }
697
698 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicAccess& access, Node* val, const Type* value_type) const {
699 GraphKit* kit = access.kit();
700 if (access.is_oop()) {
701 val = shenandoah_storeval_barrier(kit, val);
702 }
703 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
704 if (access.is_oop()) {
705 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, result));
706 shenandoah_write_barrier_pre(kit, false /* do_load */,
707 NULL, NULL, max_juint, NULL, NULL,
708 result /* pre_val */, T_OBJECT);
709 }
710 return result;
711 }
712
713 // Support for GC barriers emitted during parsing
714 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
715 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) return true;
716 if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) {
717 return false;
718 }
719 CallLeafNode *call = node->as_CallLeaf();
720 if (call->_name == NULL) {
721 return false;
722 }
723
724 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 ||
725 strcmp(call->_name, "shenandoah_cas_obj") == 0 ||
726 strcmp(call->_name, "shenandoah_wb_pre") == 0;
727 }
728
729 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
730 if (c == NULL) {
731 return c;
732 }
733 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
734 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
735 }
736 if (c->Opcode() == Op_ShenandoahEnqueueBarrier) {
737 c = c->in(1);
738 }
739 return c;
740 }
741
742 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
743 return !ShenandoahBarrierC2Support::expand(C, igvn);
744 }
745
746 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
747 if (mode == LoopOptsShenandoahExpand) {
748 assert(UseShenandoahGC, "only for shenandoah");
749 ShenandoahBarrierC2Support::pin_and_expand(phase);
750 return true;
751 } else if (mode == LoopOptsShenandoahPostExpand) {
752 assert(UseShenandoahGC, "only for shenandoah");
753 visited.Clear();
754 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
755 return true;
756 }
757 return false;
758 }
759
760 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(BasicType type) const {
761 return false;
762 }
763
764 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
765 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
766 if (src_type->isa_instptr() != NULL) {
767 ciInstanceKlass* ik = src_type->klass()->as_instance_klass();
768 if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) {
769 if (ik->has_object_fields()) {
770 return true;
771 } else {
772 if (!src_type->klass_is_exact()) {
773 Compile::current()->dependencies()->assert_leaf_type(ik);
774 }
775 }
776 } else {
777 return true;
778 }
779 } else if (src_type->isa_aryptr()) {
780 BasicType src_elem = src_type->klass()->as_array_klass()->element_type()->basic_type();
781 if (src_elem == T_OBJECT || src_elem == T_ARRAY) {
782 return true;
783 }
784 } else {
785 return true;
786 }
787 return false;
788 }
789
790 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
791 Node* ctrl = ac->in(TypeFunc::Control);
792 Node* mem = ac->in(TypeFunc::Memory);
793 Node* src = ac->in(ArrayCopyNode::Src);
794 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
795 Node* dest = ac->in(ArrayCopyNode::Dest);
796 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
797 Node* length = ac->in(ArrayCopyNode::Length);
798 assert (src_offset == NULL && dest_offset == NULL, "for clone offsets should be null");
799 assert (src->is_AddP(), "for clone the src should be the interior ptr");
800 assert (dest->is_AddP(), "for clone the dst should be the interior ptr");
801
802 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
803 // Check if heap is has forwarded objects. If it does, we need to call into the special
804 // routine that would fix up source references before we can continue.
805
806 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
807 Node* region = new RegionNode(PATH_LIMIT);
808 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM);
809
810 Node* thread = phase->transform_later(new ThreadLocalNode());
811 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
812 Node* gc_state_addr = phase->transform_later(new AddPNode(phase->C->top(), thread, offset));
813
814 uint gc_state_idx = Compile::AliasIdxRaw;
815 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument
816 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
817
818 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered));
819 int flags = ShenandoahHeap::HAS_FORWARDED;
820 if (ShenandoahStoreValEnqueueBarrier) {
821 flags |= ShenandoahHeap::MARKING;
822 }
823 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(flags)));
824 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT)));
825 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne));
826
827 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If();
828 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff));
829 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff));
830
831 // Heap is stable, no need to do anything additional
832 region->init_req(_heap_stable, stable_ctrl);
833 mem_phi->init_req(_heap_stable, mem);
834
835 // Heap is unstable, call into clone barrier stub
836 Node* call = phase->make_leaf_call(unstable_ctrl, mem,
837 ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(),
838 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier),
839 "shenandoah_clone",
840 TypeRawPtr::BOTTOM,
841 src->in(AddPNode::Base));
842 call = phase->transform_later(call);
843
844 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control));
845 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory));
846 region->init_req(_heap_unstable, ctrl);
847 mem_phi->init_req(_heap_unstable, mem);
848
849 // Wire up the actual arraycopy stub now
850 ctrl = phase->transform_later(region);
851 mem = phase->transform_later(mem_phi);
852
853 const char* name = "arraycopy";
854 call = phase->make_leaf_call(ctrl, mem,
855 OptoRuntime::fast_arraycopy_Type(),
856 phase->basictype2arraycopy(T_LONG, NULL, NULL, true, name, true),
857 name, TypeRawPtr::BOTTOM,
858 src, dest, length
859 LP64_ONLY(COMMA phase->top()));
860 call = phase->transform_later(call);
861
862 // Hook up the whole thing into the graph
863 phase->igvn().replace_node(ac, call);
864 } else {
865 BarrierSetC2::clone_at_expansion(phase, ac);
866 }
867 }
868
869 // Support for macro expanded GC barriers
870 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
871 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
872 state()->add_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
873 }
874 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
875 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
876 }
877 }
878
879 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
880 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
881 state()->remove_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
882 }
883 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
884 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
885 }
886 }
887
888 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const {
889 if (is_shenandoah_wb_pre_call(n)) {
890 shenandoah_eliminate_wb_pre(n, ¯o->igvn());
891 }
892 }
893
894 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
895 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
896 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
897 c = c->unique_ctrl_out();
898 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
899 c = c->unique_ctrl_out();
900 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
901 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
902 assert(iff->is_If(), "expect test");
903 if (!is_shenandoah_marking_if(igvn, iff)) {
904 c = c->unique_ctrl_out();
905 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
906 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
907 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
908 }
909 Node* cmpx = iff->in(1)->in(1);
910 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
911 igvn->rehash_node_delayed(call);
912 call->del_req(call->req()-1);
913 }
914
915 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
916 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
917 igvn->add_users_to_worklist(node);
918 }
919 }
920
921 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful) const {
922 for (uint i = 0; i < useful.size(); i++) {
923 Node* n = useful.at(i);
924 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
925 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
926 Compile::current()->record_for_igvn(n->fast_out(i));
927 }
928 }
929 }
930 for (int i = state()->enqueue_barriers_count() - 1; i >= 0; i--) {
931 ShenandoahEnqueueBarrierNode* n = state()->enqueue_barrier(i);
932 if (!useful.member(n)) {
933 state()->remove_enqueue_barrier(n);
934 }
935 }
936 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
937 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
938 if (!useful.member(n)) {
939 state()->remove_load_reference_barrier(n);
940 }
941 }
942 }
943
944 void ShenandoahBarrierSetC2::add_users_to_worklist(Unique_Node_List* worklist) const {}
945
946 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
947 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
948 }
949
950 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
951 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
952 }
953
954 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
955 // expanded later, then now is the time to do so.
956 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
957
958 #ifdef ASSERT
959 void ShenandoahBarrierSetC2::verify_gc_barriers(bool post_parse) const {
960 if (ShenandoahVerifyOptoBarriers && !post_parse) {
961 ShenandoahBarrierC2Support::verify(Compile::current()->root());
962 }
963 }
964 #endif
965
966 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
967 if (is_shenandoah_wb_pre_call(n)) {
968 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
969 if (n->req() > cnt) {
970 Node* addp = n->in(cnt);
971 if (has_only_shenandoah_wb_pre_uses(addp)) {
972 n->del_req(cnt);
973 if (can_reshape) {
974 phase->is_IterGVN()->_worklist.push(addp);
975 }
976 return n;
977 }
978 }
979 }
980 if (n->Opcode() == Op_CmpP) {
981 Node* in1 = n->in(1);
982 Node* in2 = n->in(2);
983 if (in1->bottom_type() == TypePtr::NULL_PTR) {
984 in2 = step_over_gc_barrier(in2);
985 }
986 if (in2->bottom_type() == TypePtr::NULL_PTR) {
987 in1 = step_over_gc_barrier(in1);
988 }
989 PhaseIterGVN* igvn = phase->is_IterGVN();
990 if (in1 != n->in(1)) {
991 if (igvn != NULL) {
992 n->set_req_X(1, in1, igvn);
993 } else {
994 n->set_req(1, in1);
995 }
996 assert(in2 == n->in(2), "only one change");
997 return n;
998 }
999 if (in2 != n->in(2)) {
1000 if (igvn != NULL) {
1001 n->set_req_X(2, in2, igvn);
1002 } else {
1003 n->set_req(2, in2);
1004 }
1005 return n;
1006 }
1007 } else if (can_reshape &&
1008 n->Opcode() == Op_If &&
1009 ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1010 n->in(0) != NULL) {
1011 Node* dom = n->in(0);
1012 Node* prev_dom = n;
1013 int op = n->Opcode();
1014 int dist = 16;
1015 // Search up the dominator tree for another heap stable test
1016 while (dom->Opcode() != op || // Not same opcode?
1017 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1?
1018 prev_dom->in(0) != dom) { // One path of test does not dominate?
1019 if (dist < 0) return NULL;
1020
1021 dist--;
1022 prev_dom = dom;
1023 dom = IfNode::up_one_dom(dom);
1024 if (!dom) return NULL;
1025 }
1026
1027 // Check that we did not follow a loop back to ourselves
1028 if (n == dom) {
1029 return NULL;
1030 }
1031
1032 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN());
1033 }
1034 return NULL;
1035 }
1036
1037 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1038 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1039 Node* u = n->fast_out(i);
1040 if (!is_shenandoah_wb_pre_call(u)) {
1041 return false;
1042 }
1043 }
1044 return n->outcnt() > 0;
1045 }
1046
1047 Node* ShenandoahBarrierSetC2::arraycopy_load_reference_barrier(PhaseGVN *phase, Node* v) {
1048 if (ShenandoahLoadRefBarrier) {
1049 return phase->transform(new ShenandoahLoadReferenceBarrierNode(NULL, v));
1050 }
1051 if (ShenandoahStoreValEnqueueBarrier) {
1052 return phase->transform(new ShenandoahEnqueueBarrierNode(v));
1053 }
1054 return v;
1055 }
1056