1 /*
2 * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "gc/shenandoah/shenandoahBarrierSet.hpp"
28 #include "gc/shenandoah/shenandoahForwarding.hpp"
29 #include "gc/shenandoah/shenandoahHeap.hpp"
30 #include "gc/shenandoah/shenandoahRuntime.hpp"
31 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
32 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
33 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
34 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
35 #include "opto/arraycopynode.hpp"
36 #include "opto/escape.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/idealKit.hpp"
39 #include "opto/macro.hpp"
40 #include "opto/movenode.hpp"
41 #include "opto/narrowptrnode.hpp"
42 #include "opto/rootnode.hpp"
43 #include "opto/runtime.hpp"
44
45 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() {
46 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2());
47 }
48
49 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena)
50 : _enqueue_barriers(new (comp_arena) GrowableArray<ShenandoahEnqueueBarrierNode*>(comp_arena, 8, 0, NULL)),
51 _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8, 0, NULL)) {
52 }
53
54 int ShenandoahBarrierSetC2State::enqueue_barriers_count() const {
55 return _enqueue_barriers->length();
56 }
57
58 ShenandoahEnqueueBarrierNode* ShenandoahBarrierSetC2State::enqueue_barrier(int idx) const {
59 return _enqueue_barriers->at(idx);
60 }
61
62 void ShenandoahBarrierSetC2State::add_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) {
63 assert(!_enqueue_barriers->contains(n), "duplicate entry in barrier list");
64 _enqueue_barriers->append(n);
65 }
66
67 void ShenandoahBarrierSetC2State::remove_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) {
68 _enqueue_barriers->remove_if_existing(n);
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] = TypeOopPtr::BOTTOM; // 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] = TypeOopPtr::BOTTOM;
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
527 if (ShenandoahStoreValEnqueueBarrier) {
528 Node* enqueue = gvn.transform(new ShenandoahEnqueueBarrierNode(val.node()));
529 val.set_node(enqueue);
530 }
531 }
532 return BarrierSetC2::store_at_resolved(access, val);
533 }
534
535 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
536 // 1: non-reference load, no additional barrier is needed
537 if (!access.is_oop()) {
538 return BarrierSetC2::load_at_resolved(access, val_type);;
539 }
540
541 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
542 DecoratorSet decorators = access.decorators();
543 BasicType type = access.type();
544
545 // 2: apply LRB if needed
546 if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
547 load = new ShenandoahLoadReferenceBarrierNode(NULL,
548 load,
549 ShenandoahBarrierSet::use_load_reference_barrier_native(decorators, type));
550 if (access.is_parse_access()) {
551 load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load);
552 } else {
553 load = static_cast<C2OptAccess &>(access).gvn().transform(load);
554 }
555 }
556
557 // 3: apply keep-alive barrier if needed
558 if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
559 Node* top = Compile::current()->top();
560 Node* adr = access.addr().node();
561 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
562 Node* obj = access.base();
563
564 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
565 bool on_weak_ref = (decorators & (ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF)) != 0;
566 bool keep_alive = (decorators & AS_NO_KEEPALIVE) == 0;
567
568 // If we are reading the value of the referent field of a Reference
569 // object (either by using Unsafe directly or through reflection)
570 // then, if SATB is enabled, we need to record the referent in an
571 // SATB log buffer using the pre-barrier mechanism.
572 // Also we need to add memory barrier to prevent commoning reads
573 // from this field across safepoint since GC can change its value.
574 if (!on_weak_ref || (unknown && (offset == top || obj == top)) || !keep_alive) {
575 return load;
576 }
577
578 assert(access.is_parse_access(), "entry not supported at optimization time");
579 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
580 GraphKit* kit = parse_access.kit();
581 bool mismatched = (decorators & C2_MISMATCHED) != 0;
582 bool is_unordered = (decorators & MO_UNORDERED) != 0;
583 bool in_native = (decorators & IN_NATIVE) != 0;
584 bool need_cpu_mem_bar = !is_unordered || mismatched || in_native;
585
586 if (on_weak_ref) {
587 // Use the pre-barrier to record the value in the referent field
588 satb_write_barrier_pre(kit, false /* do_load */,
589 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
590 load /* pre_val */, T_OBJECT);
591 // Add memory barrier to prevent commoning reads from this field
592 // across safepoint since GC can change its value.
593 kit->insert_mem_bar(Op_MemBarCPUOrder);
594 } else if (unknown) {
595 // We do not require a mem bar inside pre_barrier if need_mem_bar
596 // is set: the barriers would be emitted by us.
597 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
598 }
599 }
600
601 return load;
602 }
603
604 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
605 Node* new_val, const Type* value_type) const {
606 GraphKit* kit = access.kit();
607 if (access.is_oop()) {
608 new_val = shenandoah_storeval_barrier(kit, new_val);
609 shenandoah_write_barrier_pre(kit, false /* do_load */,
610 NULL, NULL, max_juint, NULL, NULL,
611 expected_val /* pre_val */, T_OBJECT);
612
613 MemNode::MemOrd mo = access.mem_node_mo();
614 Node* mem = access.memory();
615 Node* adr = access.addr().node();
616 const TypePtr* adr_type = access.addr().type();
617 Node* load_store = NULL;
618
619 #ifdef _LP64
620 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
621 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
622 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
623 if (ShenandoahCASBarrier) {
624 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
625 } else {
626 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
627 }
628 } else
629 #endif
630 {
631 if (ShenandoahCASBarrier) {
632 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
633 } else {
634 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
635 }
636 }
637
638 access.set_raw_access(load_store);
639 pin_atomic_op(access);
640
641 #ifdef _LP64
642 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
643 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
644 }
645 #endif
646 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, load_store, false));
647 return load_store;
648 }
649 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
650 }
651
652 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
653 Node* new_val, const Type* value_type) const {
654 GraphKit* kit = access.kit();
655 if (access.is_oop()) {
656 new_val = shenandoah_storeval_barrier(kit, new_val);
657 shenandoah_write_barrier_pre(kit, false /* do_load */,
658 NULL, NULL, max_juint, NULL, NULL,
659 expected_val /* pre_val */, T_OBJECT);
660 DecoratorSet decorators = access.decorators();
661 MemNode::MemOrd mo = access.mem_node_mo();
662 Node* mem = access.memory();
663 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
664 Node* load_store = NULL;
665 Node* adr = access.addr().node();
666 #ifdef _LP64
667 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
668 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
669 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
670 if (ShenandoahCASBarrier) {
671 if (is_weak_cas) {
672 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
673 } else {
674 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
675 }
676 } else {
677 if (is_weak_cas) {
678 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
679 } else {
680 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
681 }
682 }
683 } else
684 #endif
685 {
686 if (ShenandoahCASBarrier) {
687 if (is_weak_cas) {
688 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
689 } else {
690 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
691 }
692 } else {
693 if (is_weak_cas) {
694 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
695 } else {
696 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
697 }
698 }
699 }
700 access.set_raw_access(load_store);
701 pin_atomic_op(access);
702 return load_store;
703 }
704 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
705 }
706
707 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
708 GraphKit* kit = access.kit();
709 if (access.is_oop()) {
710 val = shenandoah_storeval_barrier(kit, val);
711 }
712 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
713 if (access.is_oop()) {
714 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, result, false));
715 shenandoah_write_barrier_pre(kit, false /* do_load */,
716 NULL, NULL, max_juint, NULL, NULL,
717 result /* pre_val */, T_OBJECT);
718 }
719 return result;
720 }
721
722 // Support for GC barriers emitted during parsing
723 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
724 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) return true;
725 if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) {
726 return false;
727 }
728 CallLeafNode *call = node->as_CallLeaf();
729 if (call->_name == NULL) {
730 return false;
731 }
732
733 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 ||
734 strcmp(call->_name, "shenandoah_cas_obj") == 0 ||
735 strcmp(call->_name, "shenandoah_wb_pre") == 0;
736 }
737
738 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
739 if (c == NULL) {
740 return c;
741 }
742 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
743 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
744 }
745 if (c->Opcode() == Op_ShenandoahEnqueueBarrier) {
746 c = c->in(1);
747 }
748 return c;
749 }
750
751 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
752 return !ShenandoahBarrierC2Support::expand(C, igvn);
753 }
754
755 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
756 if (mode == LoopOptsShenandoahExpand) {
757 assert(UseShenandoahGC, "only for shenandoah");
758 ShenandoahBarrierC2Support::pin_and_expand(phase);
759 return true;
760 } else if (mode == LoopOptsShenandoahPostExpand) {
761 assert(UseShenandoahGC, "only for shenandoah");
762 visited.clear();
763 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase);
764 return true;
765 }
766 return false;
767 }
768
769 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, ArrayCopyPhase phase) const {
770 bool is_oop = is_reference_type(type);
771 if (!is_oop) {
772 return false;
773 }
774 if (ShenandoahSATBBarrier && tightly_coupled_alloc) {
775 if (phase == Optimization) {
776 return false;
777 }
778 return !is_clone;
779 }
780 if (phase == Optimization) {
781 return !ShenandoahStoreValEnqueueBarrier;
782 }
783 return true;
784 }
785
786 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
787 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
788 if (src_type->isa_instptr() != NULL) {
789 ciInstanceKlass* ik = src_type->klass()->as_instance_klass();
790 if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) {
791 if (ik->has_object_fields()) {
792 return true;
793 } else {
794 if (!src_type->klass_is_exact()) {
795 Compile::current()->dependencies()->assert_leaf_type(ik);
796 }
797 }
798 } else {
799 return true;
800 }
801 } else if (src_type->isa_aryptr()) {
802 BasicType src_elem = src_type->klass()->as_array_klass()->element_type()->basic_type();
803 if (is_reference_type(src_elem)) {
804 return true;
805 }
806 } else {
807 return true;
808 }
809 return false;
810 }
811
812 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
813 Node* ctrl = ac->in(TypeFunc::Control);
814 Node* mem = ac->in(TypeFunc::Memory);
815 Node* src_base = ac->in(ArrayCopyNode::Src);
816 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
817 Node* dest_base = ac->in(ArrayCopyNode::Dest);
818 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
819 Node* length = ac->in(ArrayCopyNode::Length);
820
821 Node* src = phase->basic_plus_adr(src_base, src_offset);
822 Node* dest = phase->basic_plus_adr(dest_base, dest_offset);
823
824 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
825 // Check if heap is has forwarded objects. If it does, we need to call into the special
826 // routine that would fix up source references before we can continue.
827
828 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
829 Node* region = new RegionNode(PATH_LIMIT);
830 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM);
831
832 Node* thread = phase->transform_later(new ThreadLocalNode());
833 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
834 Node* gc_state_addr = phase->transform_later(new AddPNode(phase->C->top(), thread, offset));
835
836 uint gc_state_idx = Compile::AliasIdxRaw;
837 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument
838 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
839
840 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered));
841 int flags = ShenandoahHeap::HAS_FORWARDED;
842 if (ShenandoahStoreValEnqueueBarrier) {
843 flags |= ShenandoahHeap::MARKING;
844 }
845 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(flags)));
846 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT)));
847 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne));
848
849 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If();
850 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff));
851 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff));
852
853 // Heap is stable, no need to do anything additional
854 region->init_req(_heap_stable, stable_ctrl);
855 mem_phi->init_req(_heap_stable, mem);
856
857 // Heap is unstable, call into clone barrier stub
858 Node* call = phase->make_leaf_call(unstable_ctrl, mem,
859 ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(),
860 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier),
861 "shenandoah_clone",
862 TypeRawPtr::BOTTOM,
863 src_base);
864 call = phase->transform_later(call);
865
866 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control));
867 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory));
868 region->init_req(_heap_unstable, ctrl);
869 mem_phi->init_req(_heap_unstable, mem);
870
871 // Wire up the actual arraycopy stub now
872 ctrl = phase->transform_later(region);
873 mem = phase->transform_later(mem_phi);
874
875 const char* name = "arraycopy";
876 call = phase->make_leaf_call(ctrl, mem,
877 OptoRuntime::fast_arraycopy_Type(),
878 phase->basictype2arraycopy(T_LONG, NULL, NULL, true, name, true),
879 name, TypeRawPtr::BOTTOM,
880 src, dest, length
881 LP64_ONLY(COMMA phase->top()));
882 call = phase->transform_later(call);
883
884 // Hook up the whole thing into the graph
885 phase->igvn().replace_node(ac, call);
886 } else {
887 BarrierSetC2::clone_at_expansion(phase, ac);
888 }
889 }
890
891
892 // Support for macro expanded GC barriers
893 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
894 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
895 state()->add_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
896 }
897 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
898 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
899 }
900 }
901
902 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
903 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) {
904 state()->remove_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node);
905 }
906 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
907 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
908 }
909 }
910
911 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const {
912 if (is_shenandoah_wb_pre_call(n)) {
913 shenandoah_eliminate_wb_pre(n, ¯o->igvn());
914 }
915 }
916
917 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
918 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
919 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
920 c = c->unique_ctrl_out();
921 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
922 c = c->unique_ctrl_out();
923 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
924 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
925 assert(iff->is_If(), "expect test");
926 if (!is_shenandoah_marking_if(igvn, iff)) {
927 c = c->unique_ctrl_out();
928 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
929 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
930 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
931 }
932 Node* cmpx = iff->in(1)->in(1);
933 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
934 igvn->rehash_node_delayed(call);
935 call->del_req(call->req()-1);
936 }
937
938 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
939 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
940 igvn->add_users_to_worklist(node);
941 }
942 }
943
944 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
945 for (uint i = 0; i < useful.size(); i++) {
946 Node* n = useful.at(i);
947 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
948 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
949 C->record_for_igvn(n->fast_out(i));
950 }
951 }
952 }
953 for (int i = state()->enqueue_barriers_count() - 1; i >= 0; i--) {
954 ShenandoahEnqueueBarrierNode* n = state()->enqueue_barrier(i);
955 if (!useful.member(n)) {
956 state()->remove_enqueue_barrier(n);
957 }
958 }
959 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
960 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
961 if (!useful.member(n)) {
962 state()->remove_load_reference_barrier(n);
963 }
964 }
965 }
966
967 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
968 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
969 }
970
971 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
972 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
973 }
974
975 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
976 // expanded later, then now is the time to do so.
977 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
978
979 #ifdef ASSERT
980 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
981 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) {
982 ShenandoahBarrierC2Support::verify(Compile::current()->root());
983 } else if (phase == BarrierSetC2::BeforeCodeGen) {
984 // Verify G1 pre-barriers
985 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset());
986
987 Unique_Node_List visited;
988 Node_List worklist;
989 // We're going to walk control flow backwards starting from the Root
990 worklist.push(compile->root());
991 while (worklist.size() > 0) {
992 Node *x = worklist.pop();
993 if (x == NULL || x == compile->top()) continue;
994 if (visited.member(x)) {
995 continue;
996 } else {
997 visited.push(x);
998 }
999
1000 if (x->is_Region()) {
1001 for (uint i = 1; i < x->req(); i++) {
1002 worklist.push(x->in(i));
1003 }
1004 } else {
1005 worklist.push(x->in(0));
1006 // We are looking for the pattern:
1007 // /->ThreadLocal
1008 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
1009 // \->ConI(0)
1010 // We want to verify that the If and the LoadB have the same control
1011 // See GraphKit::g1_write_barrier_pre()
1012 if (x->is_If()) {
1013 IfNode *iff = x->as_If();
1014 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
1015 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
1016 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
1017 && cmp->in(1)->is_Load()) {
1018 LoadNode *load = cmp->in(1)->as_Load();
1019 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
1020 && load->in(2)->in(3)->is_Con()
1021 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
1022
1023 Node *if_ctrl = iff->in(0);
1024 Node *load_ctrl = load->in(0);
1025
1026 if (if_ctrl != load_ctrl) {
1027 // Skip possible CProj->NeverBranch in infinite loops
1028 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
1029 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
1030 if_ctrl = if_ctrl->in(0)->in(0);
1031 }
1032 }
1033 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
1034 }
1035 }
1036 }
1037 }
1038 }
1039 }
1040 }
1041 }
1042 #endif
1043
1044 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
1045 if (is_shenandoah_wb_pre_call(n)) {
1046 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1047 if (n->req() > cnt) {
1048 Node* addp = n->in(cnt);
1049 if (has_only_shenandoah_wb_pre_uses(addp)) {
1050 n->del_req(cnt);
1051 if (can_reshape) {
1052 phase->is_IterGVN()->_worklist.push(addp);
1053 }
1054 return n;
1055 }
1056 }
1057 }
1058 if (n->Opcode() == Op_CmpP) {
1059 Node* in1 = n->in(1);
1060 Node* in2 = n->in(2);
1061 if (in1->bottom_type() == TypePtr::NULL_PTR) {
1062 in2 = step_over_gc_barrier(in2);
1063 }
1064 if (in2->bottom_type() == TypePtr::NULL_PTR) {
1065 in1 = step_over_gc_barrier(in1);
1066 }
1067 PhaseIterGVN* igvn = phase->is_IterGVN();
1068 if (in1 != n->in(1)) {
1069 if (igvn != NULL) {
1070 n->set_req_X(1, in1, igvn);
1071 } else {
1072 n->set_req(1, in1);
1073 }
1074 assert(in2 == n->in(2), "only one change");
1075 return n;
1076 }
1077 if (in2 != n->in(2)) {
1078 if (igvn != NULL) {
1079 n->set_req_X(2, in2, igvn);
1080 } else {
1081 n->set_req(2, in2);
1082 }
1083 return n;
1084 }
1085 } else if (can_reshape &&
1086 n->Opcode() == Op_If &&
1087 ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1088 n->in(0) != NULL) {
1089 Node* dom = n->in(0);
1090 Node* prev_dom = n;
1091 int op = n->Opcode();
1092 int dist = 16;
1093 // Search up the dominator tree for another heap stable test
1094 while (dom->Opcode() != op || // Not same opcode?
1095 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1?
1096 prev_dom->in(0) != dom) { // One path of test does not dominate?
1097 if (dist < 0) return NULL;
1098
1099 dist--;
1100 prev_dom = dom;
1101 dom = IfNode::up_one_dom(dom);
1102 if (!dom) return NULL;
1103 }
1104
1105 // Check that we did not follow a loop back to ourselves
1106 if (n == dom) {
1107 return NULL;
1108 }
1109
1110 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN());
1111 }
1112
1113 return NULL;
1114 }
1115
1116 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1117 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1118 Node* u = n->fast_out(i);
1119 if (!is_shenandoah_wb_pre_call(u)) {
1120 return false;
1121 }
1122 }
1123 return n->outcnt() > 0;
1124 }
1125
1126 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode) const {
1127 switch (opcode) {
1128 case Op_CallLeaf:
1129 case Op_CallLeafNoFP: {
1130 assert (n->is_Call(), "");
1131 CallNode *call = n->as_Call();
1132 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1133 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1134 if (call->req() > cnt) {
1135 assert(call->req() == cnt + 1, "only one extra input");
1136 Node *addp = call->in(cnt);
1137 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1138 call->del_req(cnt);
1139 }
1140 }
1141 return false;
1142 }
1143 case Op_ShenandoahCompareAndSwapP:
1144 case Op_ShenandoahCompareAndSwapN:
1145 case Op_ShenandoahWeakCompareAndSwapN:
1146 case Op_ShenandoahWeakCompareAndSwapP:
1147 case Op_ShenandoahCompareAndExchangeP:
1148 case Op_ShenandoahCompareAndExchangeN:
1149 #ifdef ASSERT
1150 if( VerifyOptoOopOffsets ) {
1151 MemNode* mem = n->as_Mem();
1152 // Check to see if address types have grounded out somehow.
1153 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
1154 ciInstanceKlass *k = tp->klass()->as_instance_klass();
1155 bool oop_offset_is_sane = k->contains_field_offset(tp->offset());
1156 assert( !tp || oop_offset_is_sane, "" );
1157 }
1158 #endif
1159 return true;
1160 case Op_ShenandoahLoadReferenceBarrier:
1161 assert(false, "should have been expanded already");
1162 return true;
1163 default:
1164 return false;
1165 }
1166 }
1167
1168 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1169 switch (opcode) {
1170 case Op_ShenandoahCompareAndExchangeP:
1171 case Op_ShenandoahCompareAndExchangeN:
1172 conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1173 // fallthrough
1174 case Op_ShenandoahWeakCompareAndSwapP:
1175 case Op_ShenandoahWeakCompareAndSwapN:
1176 case Op_ShenandoahCompareAndSwapP:
1177 case Op_ShenandoahCompareAndSwapN:
1178 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1179 return true;
1180 case Op_StoreP: {
1181 Node* adr = n->in(MemNode::Address);
1182 const Type* adr_type = gvn->type(adr);
1183 // Pointer stores in G1 barriers looks like unsafe access.
1184 // Ignore such stores to be able scalar replace non-escaping
1185 // allocations.
1186 if (adr_type->isa_rawptr() && adr->is_AddP()) {
1187 Node* base = conn_graph->get_addp_base(adr);
1188 if (base->Opcode() == Op_LoadP &&
1189 base->in(MemNode::Address)->is_AddP()) {
1190 adr = base->in(MemNode::Address);
1191 Node* tls = conn_graph->get_addp_base(adr);
1192 if (tls->Opcode() == Op_ThreadLocal) {
1193 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1194 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1195 if (offs == buf_offset) {
1196 return true; // Pre barrier previous oop value store.
1197 }
1198 }
1199 }
1200 }
1201 return false;
1202 }
1203 case Op_ShenandoahEnqueueBarrier:
1204 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1205 break;
1206 case Op_ShenandoahLoadReferenceBarrier:
1207 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist);
1208 return true;
1209 default:
1210 // Nothing
1211 break;
1212 }
1213 return false;
1214 }
1215
1216 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1217 switch (opcode) {
1218 case Op_ShenandoahCompareAndExchangeP:
1219 case Op_ShenandoahCompareAndExchangeN: {
1220 Node *adr = n->in(MemNode::Address);
1221 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
1222 // fallthrough
1223 }
1224 case Op_ShenandoahCompareAndSwapP:
1225 case Op_ShenandoahCompareAndSwapN:
1226 case Op_ShenandoahWeakCompareAndSwapP:
1227 case Op_ShenandoahWeakCompareAndSwapN:
1228 return conn_graph->add_final_edges_unsafe_access(n, opcode);
1229 case Op_ShenandoahEnqueueBarrier:
1230 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), NULL);
1231 return true;
1232 case Op_ShenandoahLoadReferenceBarrier:
1233 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), NULL);
1234 return true;
1235 default:
1236 // Nothing
1237 break;
1238 }
1239 return false;
1240 }
1241
1242 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1243 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1244 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1245
1246 }
1247
1248 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1249 switch (opcode) {
1250 case Op_ShenandoahCompareAndExchangeP:
1251 case Op_ShenandoahCompareAndExchangeN:
1252 case Op_ShenandoahWeakCompareAndSwapP:
1253 case Op_ShenandoahWeakCompareAndSwapN:
1254 case Op_ShenandoahCompareAndSwapP:
1255 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree
1256 Node* newval = n->in(MemNode::ValueIn);
1257 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1258 Node* pair = new BinaryNode(oldval, newval);
1259 n->set_req(MemNode::ValueIn,pair);
1260 n->del_req(LoadStoreConditionalNode::ExpectedIn);
1261 return true;
1262 }
1263 default:
1264 break;
1265 }
1266 return false;
1267 }
1268
1269 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1270 return xop == Op_ShenandoahCompareAndExchangeP ||
1271 xop == Op_ShenandoahCompareAndExchangeN ||
1272 xop == Op_ShenandoahWeakCompareAndSwapP ||
1273 xop == Op_ShenandoahWeakCompareAndSwapN ||
1274 xop == Op_ShenandoahCompareAndSwapN ||
1275 xop == Op_ShenandoahCompareAndSwapP;
1276 }