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 Shenandoah 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 }
308
309 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseTransform *phase, Node* n) {
310 if (n->Opcode() != Op_If) {
311 return false;
312 }
313
314 Node* bol = n->in(1);
315 assert(bol->is_Bool(), "");
316 Node* cmpx = bol->in(1);
317 if (bol->as_Bool()->_test._test == BoolTest::ne &&
318 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
319 is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
320 cmpx->in(1)->in(2)->is_Con() &&
321 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
322 return true;
323 }
324
325 return false;
326 }
327
328 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
329 if (!n->is_Load()) return false;
330 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
331 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
332 && n->in(2)->in(3)->is_Con()
333 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
334 }
335
336 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
337 bool do_load,
338 Node* obj,
339 Node* adr,
340 uint alias_idx,
341 Node* val,
342 const TypeOopPtr* val_type,
343 Node* pre_val,
344 BasicType bt) const {
345 if (ShenandoahSATBBarrier) {
346 IdealKit ideal(kit);
347 kit->sync_kit(ideal);
348
349 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
350
351 ideal.sync_kit(kit);
352 kit->final_sync(ideal);
353 }
354 }
355
356 Node* ShenandoahBarrierSetC2::shenandoah_enqueue_barrier(GraphKit* kit, Node* pre_val) const {
357 return kit->gvn().transform(new ShenandoahEnqueueBarrierNode(pre_val));
358 }
359
360 // Helper that guards and inserts a pre-barrier.
361 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
362 Node* pre_val, bool need_mem_bar) const {
363 // We could be accessing the referent field of a reference object. If so, when G1
364 // is enabled, we need to log the value in the referent field in an SATB buffer.
365 // This routine performs some compile time filters and generates suitable
366 // runtime filters that guard the pre-barrier code.
367 // Also add memory barrier for non volatile load from the referent field
368 // to prevent commoning of loads across safepoint.
369
370 // Some compile time checks.
371
372 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
373 const TypeX* otype = offset->find_intptr_t_type();
374 if (otype != NULL && otype->is_con() &&
375 otype->get_con() != java_lang_ref_Reference::referent_offset) {
376 // Constant offset but not the reference_offset so just return
377 return;
378 }
379
380 // We only need to generate the runtime guards for instances.
381 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
382 if (btype != NULL) {
383 if (btype->isa_aryptr()) {
384 // Array type so nothing to do
385 return;
386 }
387
388 const TypeInstPtr* itype = btype->isa_instptr();
389 if (itype != NULL) {
390 // Can the klass of base_oop be statically determined to be
391 // _not_ a sub-class of Reference and _not_ Object?
392 ciKlass* klass = itype->klass();
393 if ( klass->is_loaded() &&
394 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
395 !kit->env()->Object_klass()->is_subtype_of(klass)) {
396 return;
397 }
398 }
399 }
400
401 // The compile time filters did not reject base_oop/offset so
402 // we need to generate the following runtime filters
403 //
404 // if (offset == java_lang_ref_Reference::_reference_offset) {
405 // if (instance_of(base, java.lang.ref.Reference)) {
406 // pre_barrier(_, pre_val, ...);
407 // }
408 // }
409
410 float likely = PROB_LIKELY( 0.999);
411 float unlikely = PROB_UNLIKELY(0.999);
412
413 IdealKit ideal(kit);
414
415 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
416
417 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
418 // Update graphKit memory and control from IdealKit.
419 kit->sync_kit(ideal);
420
421 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
422 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
423
424 // Update IdealKit memory and control from graphKit.
425 __ sync_kit(kit);
426
427 Node* one = __ ConI(1);
428 // is_instof == 0 if base_oop == NULL
429 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
430
431 // Update graphKit from IdeakKit.
432 kit->sync_kit(ideal);
433
434 // Use the pre-barrier to record the value in the referent field
435 satb_write_barrier_pre(kit, false /* do_load */,
436 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
437 pre_val /* pre_val */,
438 T_OBJECT);
439 if (need_mem_bar) {
440 // Add memory barrier to prevent commoning reads from this field
441 // across safepoint since GC can change its value.
442 kit->insert_mem_bar(Op_MemBarCPUOrder);
443 }
444 // Update IdealKit from graphKit.
445 __ sync_kit(kit);
446
447 } __ end_if(); // _ref_type != ref_none
448 } __ end_if(); // offset == referent_offset
449
450 // Final sync IdealKit and GraphKit.
451 kit->final_sync(ideal);
452 }
453
454 #undef __
455
456 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() {
457 const Type **fields = TypeTuple::fields(2);
458 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
459 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
460 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
461
462 // create result type (range)
463 fields = TypeTuple::fields(0);
464 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
465
466 return TypeFunc::make(domain, range);
467 }
468
469 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() {
470 const Type **fields = TypeTuple::fields(1);
471 fields[TypeFunc::Parms+0] = TypeOopPtr::NOTNULL; // src oop
472 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
473
474 // create result type (range)
475 fields = TypeTuple::fields(0);
476 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
477
478 return TypeFunc::make(domain, range);
479 }
480
481 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type() {
482 const Type **fields = TypeTuple::fields(2);
483 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
484 fields[TypeFunc::Parms+1] = TypeRawPtr::BOTTOM; // original load address
485
486 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
487
488 // create result type (range)
489 fields = TypeTuple::fields(1);
490 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL;
491 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
492
493 return TypeFunc::make(domain, range);
494 }
495
496 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
497 DecoratorSet decorators = access.decorators();
498
499 const TypePtr* adr_type = access.addr().type();
500 Node* adr = access.addr().node();
501
502 bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
503 bool on_heap = (decorators & IN_HEAP) != 0;
504
505 if (!access.is_oop() || (!on_heap && !anonymous)) {
506 return BarrierSetC2::store_at_resolved(access, val);
507 }
508
509 GraphKit* kit = access.kit();
510
511 uint adr_idx = kit->C->get_alias_index(adr_type);
512 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
513 Node* value = val.node();
514 value = shenandoah_storeval_barrier(kit, value);
515 val.set_node(value);
516 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
517 static_cast<const TypeOopPtr*>(val.type()), NULL /* pre_val */, access.type());
518 return BarrierSetC2::store_at_resolved(access, val);
519 }
520
521 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
522 DecoratorSet decorators = access.decorators();
523
524 Node* adr = access.addr().node();
525 Node* obj = access.base();
526
527 bool mismatched = (decorators & C2_MISMATCHED) != 0;
528 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
529 bool on_heap = (decorators & IN_HEAP) != 0;
530 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
531 bool is_unordered = (decorators & MO_UNORDERED) != 0;
532 bool need_cpu_mem_bar = !is_unordered || mismatched || !on_heap;
533
534 Node* top = Compile::current()->top();
535
536 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
537 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
538
539 if (access.is_oop()) {
540 if (ShenandoahLoadRefBarrier) {
541 load = new ShenandoahLoadReferenceBarrierNode(NULL, load);
542 load = access.kit()->gvn().transform(load);
543 }
544 }
545
546 // If we are reading the value of the referent field of a Reference
547 // object (either by using Unsafe directly or through reflection)
548 // then, if SATB is enabled, we need to record the referent in an
549 // SATB log buffer using the pre-barrier mechanism.
550 // Also we need to add memory barrier to prevent commoning reads
551 // from this field across safepoint since GC can change its value.
552 bool need_read_barrier = ShenandoahKeepAliveBarrier &&
553 (on_heap && (on_weak || (unknown && offset != top && obj != top)));
554
555 if (!access.is_oop() || !need_read_barrier) {
556 return load;
557 }
558
559 GraphKit* kit = access.kit();
560
561 if (on_weak) {
562 // Use the pre-barrier to record the value in the referent field
563 satb_write_barrier_pre(kit, false /* do_load */,
564 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
565 load /* pre_val */, T_OBJECT);
566 // Add memory barrier to prevent commoning reads from this field
567 // across safepoint since GC can change its value.
568 kit->insert_mem_bar(Op_MemBarCPUOrder);
569 } else if (unknown) {
570 // We do not require a mem bar inside pre_barrier if need_mem_bar
571 // is set: the barriers would be emitted by us.
572 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
573 }
574
575 return load;
576 }
577
578 static void pin_atomic_op(C2AtomicAccess& access) {
579 if (!access.needs_pinning()) {
580 return;
581 }
582 // SCMemProjNodes represent the memory state of a LoadStore. Their
583 // main role is to prevent LoadStore nodes from being optimized away
584 // when their results aren't used.
585 GraphKit* kit = access.kit();
586 Node* load_store = access.raw_access();
587 assert(load_store != NULL, "must pin atomic op");
588 Node* proj = kit->gvn().transform(new SCMemProjNode(load_store));
589 kit->set_memory(proj, access.alias_idx());
590 }
591
592 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicAccess& access, Node* expected_val,
593 Node* new_val, const Type* value_type) const {
594 GraphKit* kit = access.kit();
595 if (access.is_oop()) {
596 new_val = shenandoah_storeval_barrier(kit, new_val);
597 shenandoah_write_barrier_pre(kit, false /* do_load */,
598 NULL, NULL, max_juint, NULL, NULL,
599 expected_val /* pre_val */, T_OBJECT);
600
601 MemNode::MemOrd mo = access.mem_node_mo();
602 Node* mem = access.memory();
603 Node* adr = access.addr().node();
604 const TypePtr* adr_type = access.addr().type();
605 Node* load_store = NULL;
606
607 #ifdef _LP64
608 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
609 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
610 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
611 if (ShenandoahCASBarrier) {
612 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
613 } else {
614 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
615 }
616 } else
617 #endif
618 {
619 if (ShenandoahCASBarrier) {
620 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
621 } else {
622 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
623 }
624 }
625
626 access.set_raw_access(load_store);
627 pin_atomic_op(access);
628
629 #ifdef _LP64
630 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
631 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
632 }
633 #endif
634 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, load_store));
635 return load_store;
636 }
637 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
638 }
639
640 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicAccess& access, Node* expected_val,
641 Node* new_val, const Type* value_type) const {
642 GraphKit* kit = access.kit();
643 if (access.is_oop()) {
644 new_val = shenandoah_storeval_barrier(kit, new_val);
645 shenandoah_write_barrier_pre(kit, false /* do_load */,
646 NULL, NULL, max_juint, NULL, NULL,
647 expected_val /* pre_val */, T_OBJECT);
648 DecoratorSet decorators = access.decorators();
649 MemNode::MemOrd mo = access.mem_node_mo();
650 Node* mem = access.memory();
651 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
652 Node* load_store = NULL;
653 Node* adr = access.addr().node();
654 #ifdef _LP64
655 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
656 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
657 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
658 if (ShenandoahCASBarrier) {
659 if (is_weak_cas) {
660 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
661 } else {
662 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
663 }
664 } else {
665 if (is_weak_cas) {
666 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
667 } else {
668 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
669 }
670 }
671 } else
672 #endif
673 {
674 if (ShenandoahCASBarrier) {
675 if (is_weak_cas) {
676 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
677 } else {
678 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
679 }
680 } else {
681 if (is_weak_cas) {
682 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
683 } else {
684 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
685 }
686 }
687 }
688 access.set_raw_access(load_store);
689 pin_atomic_op(access);
690 return load_store;
691 }
692 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
693 }
694
695 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicAccess& access, Node* val, const Type* value_type) const {
696 GraphKit* kit = access.kit();
697 if (access.is_oop()) {
698 val = shenandoah_storeval_barrier(kit, val);
699 }
700 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
701 if (access.is_oop()) {
702 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, result));
703 shenandoah_write_barrier_pre(kit, false /* do_load */,
704 NULL, NULL, max_juint, NULL, NULL,
705 result /* pre_val */, T_OBJECT);
706 }
707 return result;
708 }
709
710 // Support for GC barriers emitted during parsing
711 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
712 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) return true;
713 if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) {
714 return false;
715 }
716 CallLeafNode *call = node->as_CallLeaf();
717 if (call->_name == NULL) {
718 return false;
719 }
720
721 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 ||
722 strcmp(call->_name, "shenandoah_cas_obj") == 0 ||
723 strcmp(call->_name, "shenandoah_wb_pre") == 0;
724 }
725
726 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
727 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
728 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
729 }
730 if (c->Opcode() == Op_ShenandoahEnqueueBarrier) {
731 c = c->in(1);
732 }
733 return c;
734 }
735
736 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
737 return !ShenandoahBarrierC2Support::expand(C, igvn);
738 }
739
740 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
741 if (mode == LoopOptsShenandoahExpand) {
742 assert(UseShenandoahGC, "only for shenandoah");
743 ShenandoahBarrierC2Support::pin_and_expand(phase);
744 return true;
745 } else if (mode == LoopOptsShenandoahPostExpand) {
746 assert(UseShenandoahGC, "only for shenandoah");
747 visited.Clear();
748 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
749 return true;
750 }
751 return false;
752 }
753
754 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(BasicType type) const {
755 return false;
756 }
757
758 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
759 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
760 if (src_type->isa_instptr() != NULL) {
761 ciInstanceKlass* ik = src_type->klass()->as_instance_klass();
762 if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) {
763 if (ik->has_object_fields()) {
764 return true;
765 } else {
766 if (!src_type->klass_is_exact()) {
767 Compile::current()->dependencies()->assert_leaf_type(ik);
768 }
769 }
770 } else {
771 return true;
772 }
773 } else if (src_type->isa_aryptr()) {
774 BasicType src_elem = src_type->klass()->as_array_klass()->element_type()->basic_type();
775 if (src_elem == T_OBJECT || src_elem == T_ARRAY) {
776 return true;
777 }
778 } else {
779 return true;
780 }
781 return false;
782 }
783
784 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
785 Node* ctrl = ac->in(TypeFunc::Control);
786 Node* mem = ac->in(TypeFunc::Memory);
787 Node* src = ac->in(ArrayCopyNode::Src);
788 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
789 Node* dest = ac->in(ArrayCopyNode::Dest);
790 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
791 Node* length = ac->in(ArrayCopyNode::Length);
792 assert (src_offset == NULL && dest_offset == NULL, "for clone offsets should be null");
793 assert (src->is_AddP(), "for clone the src should be the interior ptr");
794 assert (dest->is_AddP(), "for clone the dst should be the interior ptr");
795
796 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
797 // Check if heap is has forwarded objects. If it does, we need to call into the special
798 // routine that would fix up source references before we can continue.
799
800 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
801 Node* region = new RegionNode(PATH_LIMIT);
802 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM);
803
804 Node* thread = phase->transform_later(new ThreadLocalNode());
805 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
806 Node* gc_state_addr = phase->transform_later(new AddPNode(phase->C->top(), thread, offset));
807
808 uint gc_state_idx = Compile::AliasIdxRaw;
809 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument
810 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
811
812 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered));
813 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED)));
814 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT)));
815 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne));
816
817 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If();
818 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff));
819 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff));
820
821 // Heap is stable, no need to do anything additional
822 region->init_req(_heap_stable, stable_ctrl);
823 mem_phi->init_req(_heap_stable, mem);
824
825 // Heap is unstable, call into clone barrier stub
826 Node* call = phase->make_leaf_call(unstable_ctrl, mem,
827 ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(),
828 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier),
829 "shenandoah_clone",
830 TypeRawPtr::BOTTOM,
831 src->in(AddPNode::Base));
832 call = phase->transform_later(call);
833
834 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control));
835 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory));
836 region->init_req(_heap_unstable, ctrl);
837 mem_phi->init_req(_heap_unstable, mem);
838
839 // Wire up the actual arraycopy stub now
840 ctrl = phase->transform_later(region);
841 mem = phase->transform_later(mem_phi);
842
843 const char* name = "arraycopy";
844 call = phase->make_leaf_call(ctrl, mem,
845 OptoRuntime::fast_arraycopy_Type(),
846 phase->basictype2arraycopy(T_LONG, NULL, NULL, true, name, true),
847 name, TypeRawPtr::BOTTOM,
848 src, dest, length
849 LP64_ONLY(COMMA phase->top()));
850 call = phase->transform_later(call);
851
852 // Hook up the whole thing into the graph
853 phase->igvn().replace_node(ac, call);
854 } else {
855 BarrierSetC2::clone_at_expansion(phase, ac);
856 }
857 }
858
859 // Support for macro expanded GC barriers
860 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
861 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
862 state()->add_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
863 }
864 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
865 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
866 }
867 }
868
869 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
870 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
871 state()->remove_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
872 }
873 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
874 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
875 }
876 }
877
878 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const {
879 if (is_shenandoah_wb_pre_call(n)) {
880 shenandoah_eliminate_wb_pre(n, ¯o->igvn());
881 }
882 }
883
884 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
885 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
886 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
887 c = c->unique_ctrl_out();
888 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
889 c = c->unique_ctrl_out();
890 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
891 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
892 assert(iff->is_If(), "expect test");
893 if (!is_shenandoah_marking_if(igvn, iff)) {
894 c = c->unique_ctrl_out();
895 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
896 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
897 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
898 }
899 Node* cmpx = iff->in(1)->in(1);
900 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
901 igvn->rehash_node_delayed(call);
902 call->del_req(call->req()-1);
903 }
904
905 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(Unique_Node_List &worklist, Node* node) const {
906 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
907 worklist.push(node);
908 }
909 }
910
911 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful) const {
912 for (uint i = 0; i < useful.size(); i++) {
913 Node* n = useful.at(i);
914 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
915 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
916 Compile::current()->record_for_igvn(n->fast_out(i));
917 }
918 }
919 }
920 for (int i = state()->enqueue_barriers_count() - 1; i >= 0; i--) {
921 ShenandoahEnqueueBarrierNode* n = state()->enqueue_barrier(i);
922 if (!useful.member(n)) {
923 state()->remove_enqueue_barrier(n);
924 }
925 }
926 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
927 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
928 if (!useful.member(n)) {
929 state()->remove_load_reference_barrier(n);
930 }
931 }
932 }
933
934 void ShenandoahBarrierSetC2::add_users_to_worklist(Unique_Node_List* worklist) const {}
935
936 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
937 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
938 }
939
940 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
941 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
942 }
943
944 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
945 // expanded later, then now is the time to do so.
946 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
947
948 #ifdef ASSERT
949 void ShenandoahBarrierSetC2::verify_gc_barriers(bool post_parse) const {
950 if (ShenandoahVerifyOptoBarriers && !post_parse) {
951 ShenandoahBarrierC2Support::verify(Compile::current()->root());
952 }
953 }
954 #endif
955
956 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
957 if (is_shenandoah_wb_pre_call(n)) {
958 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
959 if (n->req() > cnt) {
960 Node* addp = n->in(cnt);
961 if (has_only_shenandoah_wb_pre_uses(addp)) {
962 n->del_req(cnt);
963 if (can_reshape) {
964 phase->is_IterGVN()->_worklist.push(addp);
965 }
966 return n;
967 }
968 }
969 }
970 if (n->Opcode() == Op_CmpP) {
971 Node* in1 = n->in(1);
972 Node* in2 = n->in(2);
973 if (in1->bottom_type() == TypePtr::NULL_PTR) {
974 in2 = step_over_gc_barrier(in2);
975 }
976 if (in2->bottom_type() == TypePtr::NULL_PTR) {
977 in1 = step_over_gc_barrier(in1);
978 }
979 PhaseIterGVN* igvn = phase->is_IterGVN();
980 if (in1 != n->in(1)) {
981 if (igvn != NULL) {
982 n->set_req_X(1, in1, igvn);
983 } else {
984 n->set_req(1, in1);
985 }
986 assert(in2 == n->in(2), "only one change");
987 return n;
988 }
989 if (in2 != n->in(2)) {
990 if (igvn != NULL) {
991 n->set_req_X(2, in2, igvn);
992 } else {
993 n->set_req(2, in2);
994 }
995 return n;
996 }
997 } else if (can_reshape &&
998 n->Opcode() == Op_If &&
999 ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1000 n->in(0) != NULL) {
1001 Node* dom = n->in(0);
1002 Node* prev_dom = n;
1003 int op = n->Opcode();
1004 int dist = 16;
1005 // Search up the dominator tree for another heap stable test
1006 while (dom->Opcode() != op || // Not same opcode?
1007 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1?
1008 prev_dom->in(0) != dom) { // One path of test does not dominate?
1009 if (dist < 0) return NULL;
1010
1011 dist--;
1012 prev_dom = dom;
1013 dom = IfNode::up_one_dom(dom);
1014 if (!dom) return NULL;
1015 }
1016
1017 // Check that we did not follow a loop back to ourselves
1018 if (n == dom) {
1019 return NULL;
1020 }
1021
1022 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN());
1023 }
1024 return NULL;
1025 }
1026
1027 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1028 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1029 Node* u = n->fast_out(i);
1030 if (!is_shenandoah_wb_pre_call(u)) {
1031 return false;
1032 }
1033 }
1034 return n->outcnt() > 0;
1035 }
1036
1037 Node* ShenandoahBarrierSetC2::arraycopy_load_reference_barrier(PhaseGVN *phase, Node* v) {
1038 if (ShenandoahLoadRefBarrier) {
1039 return phase->transform(new ShenandoahLoadReferenceBarrierNode(NULL, v));
1040 }
1041 if (ShenandoahStoreValEnqueueBarrier) {
1042 return phase->transform(new ShenandoahEnqueueBarrierNode(v));
1043 }
1044 return v;
1045 }
1046