6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "classfile/javaClasses.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
29 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
30 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
31 #include "gc/shenandoah/shenandoahBarrierSet.hpp"
32 #include "gc/shenandoah/shenandoahCardTable.hpp"
33 #include "gc/shenandoah/shenandoahForwarding.hpp"
34 #include "gc/shenandoah/shenandoahHeap.hpp"
35 #include "gc/shenandoah/shenandoahRuntime.hpp"
36 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
37 #include "opto/arraycopynode.hpp"
38 #include "opto/escape.hpp"
39 #include "opto/graphKit.hpp"
40 #include "opto/idealKit.hpp"
41 #include "opto/macro.hpp"
42 #include "opto/movenode.hpp"
43 #include "opto/narrowptrnode.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46
47 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() {
48 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2());
49 }
50
51 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena)
52 : _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8, 0, nullptr)) {
53 }
54
55 int ShenandoahBarrierSetC2State::load_reference_barriers_count() const {
56 return _load_reference_barriers->length();
57 }
58
59 ShenandoahLoadReferenceBarrierNode* ShenandoahBarrierSetC2State::load_reference_barrier(int idx) const {
60 return _load_reference_barriers->at(idx);
61 }
62
63 void ShenandoahBarrierSetC2State::add_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
64 assert(!_load_reference_barriers->contains(n), "duplicate entry in barrier list");
65 _load_reference_barriers->append(n);
66 }
67
68 void ShenandoahBarrierSetC2State::remove_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) {
69 if (_load_reference_barriers->contains(n)) {
70 _load_reference_barriers->remove(n);
71 }
72 }
73
74 #define __ kit->
75
76 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseGVN* phase, Node* adr,
77 BasicType bt, uint adr_idx) const {
78 intptr_t offset = 0;
79 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
80 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
81
82 if (offset == Type::OffsetBot) {
83 return false; // cannot unalias unless there are precise offsets
84 }
85
86 if (alloc == nullptr) {
87 return false; // No allocation found
88 }
89
90 intptr_t size_in_bytes = type2aelembytes(bt);
91
92 Node* mem = __ memory(adr_idx); // start searching here...
93
94 for (int cnt = 0; cnt < 50; cnt++) {
95
96 if (mem->is_Store()) {
97
98 Node* st_adr = mem->in(MemNode::Address);
99 intptr_t st_offset = 0;
100 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
101
102 if (st_base == nullptr) {
103 break; // inscrutable pointer
104 }
105
106 // Break we have found a store with same base and offset as ours so break
107 if (st_base == base && st_offset == offset) {
108 break;
109 }
110
111 if (st_offset != offset && st_offset != Type::OffsetBot) {
112 const int MAX_STORE = BytesPerLong;
113 if (st_offset >= offset + size_in_bytes ||
114 st_offset <= offset - MAX_STORE ||
115 st_offset <= offset - mem->as_Store()->memory_size()) {
116 // Success: The offsets are provably independent.
117 // (You may ask, why not just test st_offset != offset and be done?
118 // The answer is that stores of different sizes can co-exist
119 // in the same sequence of RawMem effects. We sometimes initialize
120 // a whole 'tile' of array elements with a single jint or jlong.)
121 mem = mem->in(MemNode::Memory);
122 continue; // advance through independent store memory
123 }
124 }
125
126 if (st_base != base
127 && MemNode::detect_ptr_independence(base, alloc, st_base,
128 AllocateNode::Ideal_allocation(st_base),
129 phase)) {
130 // Success: The bases are provably independent.
131 mem = mem->in(MemNode::Memory);
132 continue; // advance through independent store memory
133 }
134 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
135
136 InitializeNode* st_init = mem->in(0)->as_Initialize();
137 AllocateNode* st_alloc = st_init->allocation();
138
139 // Make sure that we are looking at the same allocation site.
140 // The alloc variable is guaranteed to not be null here from earlier check.
141 if (alloc == st_alloc) {
142 // Check that the initialization is storing null so that no previous store
143 // has been moved up and directly write a reference
144 Node* captured_store = st_init->find_captured_store(offset,
145 type2aelembytes(T_OBJECT),
146 phase);
147 if (captured_store == nullptr || captured_store == st_init->zero_memory()) {
148 return true;
149 }
150 }
151 }
152
153 // Unless there is an explicit 'continue', we must bail out here,
154 // because 'mem' is an inscrutable memory state (e.g., a call).
155 break;
156 }
157
158 return false;
159 }
160
161 #undef __
162 #define __ ideal.
163
164 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit,
165 bool do_load,
166 Node* obj,
167 Node* adr,
168 uint alias_idx,
169 Node* val,
170 const TypeOopPtr* val_type,
171 Node* pre_val,
172 BasicType bt) const {
173 // Some sanity checks
174 // Note: val is unused in this routine.
175
176 if (do_load) {
177 // We need to generate the load of the previous value
178 assert(adr != nullptr, "where are loading from?");
179 assert(pre_val == nullptr, "loaded already?");
180 assert(val_type != nullptr, "need a type");
181
182 if (ReduceInitialCardMarks
183 && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
184 return;
185 }
186
187 } else {
188 // In this case both val_type and alias_idx are unused.
189 assert(pre_val != nullptr, "must be loaded already");
190 // Nothing to be done if pre_val is null.
191 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
192 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
193 }
194 assert(bt == T_OBJECT, "or we shouldn't be here");
195
196 IdealKit ideal(kit, true);
197
198 Node* tls = __ thread(); // ThreadLocalStorage
199
200 Node* no_base = __ top();
201 Node* zero = __ ConI(0);
202 Node* zeroX = __ ConX(0);
203
204 float likely = PROB_LIKELY(0.999);
205 float unlikely = PROB_UNLIKELY(0.999);
206
207 // Offsets into the thread
208 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
209 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
210
211 // Now the actual pointers into the thread
212 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
213 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
214
215 // Now some of the values
216 Node* marking;
217 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())));
218 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
219 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
220 assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape");
221
222 // if (!marking)
223 __ if_then(marking, BoolTest::ne, zero, unlikely); {
224 BasicType index_bt = TypeX_X->basic_type();
225 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading Shenandoah SATBMarkQueue::_index with wrong size.");
226 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
227
228 if (do_load) {
229 // load original value
230 // alias_idx correct??
231 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
232 }
233
234 // if (pre_val != nullptr)
235 __ if_then(pre_val, BoolTest::ne, kit->null()); {
236 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
237
238 // is the queue for this thread full?
239 __ if_then(index, BoolTest::ne, zeroX, likely); {
240
241 // decrement the index
242 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
243
244 // Now get the buffer location we will log the previous value into and store it
245 Node *log_addr = __ AddP(no_base, buffer, next_index);
246 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
247 // update the index
248 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
249
250 } __ else_(); {
251
252 // logging buffer is full, call the runtime
253 const TypeFunc *tf = ShenandoahBarrierSetC2::write_barrier_pre_Type();
254 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre), "shenandoah_wb_pre", pre_val);
255 } __ end_if(); // (!index)
256 } __ end_if(); // (pre_val != nullptr)
257 } __ end_if(); // (!marking)
258
259 // Final sync IdealKit and GraphKit.
260 kit->final_sync(ideal);
261
262 if (ShenandoahSATBBarrier && adr != nullptr) {
263 Node* c = kit->control();
264 Node* call = c->in(1)->in(1)->in(1)->in(0);
265 assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected");
266 call->add_req(adr);
267 }
268 }
269
270 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) {
271 return call->is_CallLeaf() &&
272 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre);
273 }
274
275 bool ShenandoahBarrierSetC2::is_shenandoah_clone_call(Node* call) {
276 return call->is_CallLeaf() &&
277 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::clone_barrier);
278 }
279
280 bool ShenandoahBarrierSetC2::is_shenandoah_lrb_call(Node* call) {
281 if (!call->is_CallLeaf()) {
282 return false;
283 }
284
285 address entry_point = call->as_CallLeaf()->entry_point();
286 return (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong)) ||
287 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow)) ||
288 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak)) ||
289 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow)) ||
290 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom)) ||
291 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow));
292 }
293
294 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseValues* phase, Node* n) {
295 if (n->Opcode() != Op_If) {
296 return false;
297 }
298
299 Node* bol = n->in(1);
300 assert(bol->is_Bool(), "");
301 Node* cmpx = bol->in(1);
302 if (bol->as_Bool()->_test._test == BoolTest::ne &&
303 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
304 is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
305 cmpx->in(1)->in(2)->is_Con() &&
306 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
307 return true;
308 }
309
310 return false;
311 }
312
313 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
314 if (!n->is_Load()) return false;
315 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
316 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
317 && n->in(2)->in(3)->is_Con()
318 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
319 }
320
321 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
322 bool do_load,
323 Node* obj,
324 Node* adr,
325 uint alias_idx,
326 Node* val,
327 const TypeOopPtr* val_type,
328 Node* pre_val,
329 BasicType bt) const {
330 if (ShenandoahSATBBarrier) {
331 IdealKit ideal(kit);
332 kit->sync_kit(ideal);
333
334 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
335
336 ideal.sync_kit(kit);
337 kit->final_sync(ideal);
338 }
339 }
340
341 // Helper that guards and inserts a pre-barrier.
342 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
343 Node* pre_val, bool need_mem_bar) const {
344 // We could be accessing the referent field of a reference object. If so, when Shenandoah
345 // is enabled, we need to log the value in the referent field in an SATB buffer.
346 // This routine performs some compile time filters and generates suitable
347 // runtime filters that guard the pre-barrier code.
348 // Also add memory barrier for non volatile load from the referent field
349 // to prevent commoning of loads across safepoint.
350
351 // Some compile time checks.
352
353 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
354 const TypeX* otype = offset->find_intptr_t_type();
355 if (otype != nullptr && otype->is_con() &&
356 otype->get_con() != java_lang_ref_Reference::referent_offset()) {
357 // Constant offset but not the reference_offset so just return
358 return;
359 }
360
361 // We only need to generate the runtime guards for instances.
362 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
363 if (btype != nullptr) {
364 if (btype->isa_aryptr()) {
365 // Array type so nothing to do
366 return;
367 }
368
369 const TypeInstPtr* itype = btype->isa_instptr();
370 if (itype != nullptr) {
371 // Can the klass of base_oop be statically determined to be
372 // _not_ a sub-class of Reference and _not_ Object?
373 ciKlass* klass = itype->instance_klass();
374 if (klass->is_loaded() &&
375 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
376 !kit->env()->Object_klass()->is_subtype_of(klass)) {
377 return;
378 }
379 }
380 }
381
382 // The compile time filters did not reject base_oop/offset so
383 // we need to generate the following runtime filters
384 //
385 // if (offset == java_lang_ref_Reference::_reference_offset) {
386 // if (instance_of(base, java.lang.ref.Reference)) {
387 // pre_barrier(_, pre_val, ...);
388 // }
389 // }
390
391 float likely = PROB_LIKELY( 0.999);
392 float unlikely = PROB_UNLIKELY(0.999);
393
394 IdealKit ideal(kit);
395
396 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset());
397
398 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
399 // Update graphKit memory and control from IdealKit.
400 kit->sync_kit(ideal);
401
402 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
403 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
404
405 // Update IdealKit memory and control from graphKit.
406 __ sync_kit(kit);
407
408 Node* one = __ ConI(1);
409 // is_instof == 0 if base_oop == nullptr
410 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
411
412 // Update graphKit from IdeakKit.
413 kit->sync_kit(ideal);
414
415 // Use the pre-barrier to record the value in the referent field
416 satb_write_barrier_pre(kit, false /* do_load */,
417 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */,
418 pre_val /* pre_val */,
419 T_OBJECT);
420 if (need_mem_bar) {
421 // Add memory barrier to prevent commoning reads from this field
422 // across safepoint since GC can change its value.
423 kit->insert_mem_bar(Op_MemBarCPUOrder);
424 }
425 // Update IdealKit from graphKit.
426 __ sync_kit(kit);
427
428 } __ end_if(); // _ref_type != ref_none
429 } __ end_if(); // offset == referent_offset
430
431 // Final sync IdealKit and GraphKit.
432 kit->final_sync(ideal);
433 }
434
435 void ShenandoahBarrierSetC2::post_barrier(GraphKit* kit,
436 Node* ctl,
437 Node* oop_store,
438 Node* obj,
439 Node* adr,
440 uint adr_idx,
441 Node* val,
442 BasicType bt,
443 bool use_precise) const {
444 assert(ShenandoahCardBarrier, "Should have been checked by caller");
445
446 // No store check needed if we're storing a null.
447 if (val != nullptr && val->is_Con()) {
448 // must be either an oop or null
449 const Type* t = val->bottom_type();
450 if (t == TypePtr::NULL_PTR || t == Type::TOP)
451 return;
452 }
453
454 if (ReduceInitialCardMarks && obj == kit->just_allocated_object(kit->control())) {
455 // We use card marks to track old to young references in Generational Shenandoah;
456 // see flag ShenandoahCardBarrier above.
457 // Objects are always allocated in the young generation and initialized
458 // before they are promoted. There's always a safepoint (e.g. at final mark)
459 // before an object is promoted from young to old. Promotion entails dirtying of
460 // the cards backing promoted objects, so they will be guaranteed to be scanned
461 // at the next remembered set scan of the old generation.
462 // Thus, we can safely skip card-marking of initializing stores on a
463 // freshly-allocated object. If any of the assumptions above change in
464 // the future, this code will need to be re-examined; see check in
465 // ShenandoahCardBarrier::on_slowpath_allocation_exit().
466 return;
467 }
468
469 if (!use_precise) {
470 // All card marks for a (non-array) instance are in one place:
471 adr = obj;
472 }
473 // (Else it's an array (or unknown), and we want more precise card marks.)
474 assert(adr != nullptr, "");
475
476 IdealKit ideal(kit, true);
477
478 Node* tls = __ thread(); // ThreadLocalStorage
479
480 // Convert the pointer to an int prior to doing math on it
481 Node* cast = __ CastPX(__ ctrl(), adr);
482
483 Node* curr_ct_holder_offset = __ ConX(in_bytes(ShenandoahThreadLocalData::card_table_offset()));
484 Node* curr_ct_holder_addr = __ AddP(__ top(), tls, curr_ct_holder_offset);
485 Node* curr_ct_base_addr = __ load( __ ctrl(), curr_ct_holder_addr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
486
487 // Divide by card size
488 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift()) );
489
490 // Combine card table base and card offset
491 Node* card_adr = __ AddP(__ top(), curr_ct_base_addr, card_offset);
492
493 // Get the alias_index for raw card-mark memory
494 int adr_type = Compile::AliasIdxRaw;
495 Node* zero = __ ConI(0); // Dirty card value
496
497 if (UseCondCardMark) {
498 // The classic GC reference write barrier is typically implemented
499 // as a store into the global card mark table. Unfortunately
500 // unconditional stores can result in false sharing and excessive
501 // coherence traffic as well as false transactional aborts.
502 // UseCondCardMark enables MP "polite" conditional card mark
503 // stores. In theory we could relax the load from ctrl() to
504 // no_ctrl, but that doesn't buy much latitude.
505 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, T_BYTE, adr_type);
506 __ if_then(card_val, BoolTest::ne, zero);
507 }
508
509 // Smash zero into card
510 __ store(__ ctrl(), card_adr, zero, T_BYTE, adr_type, MemNode::unordered);
511
512 if (UseCondCardMark) {
513 __ end_if();
514 }
515
516 // Final sync IdealKit and GraphKit.
517 kit->final_sync(ideal);
518 }
519
520 #undef __
521
522 const TypeFunc* ShenandoahBarrierSetC2::_write_barrier_pre_Type = nullptr;
523 const TypeFunc* ShenandoahBarrierSetC2::_clone_barrier_Type = nullptr;
524 const TypeFunc* ShenandoahBarrierSetC2::_load_reference_barrier_Type = nullptr;
525
526 inline const TypeFunc* ShenandoahBarrierSetC2::write_barrier_pre_Type() {
527 assert(ShenandoahBarrierSetC2::_write_barrier_pre_Type != nullptr, "should be initialized");
528 return ShenandoahBarrierSetC2::_write_barrier_pre_Type;
529 }
530
531 inline const TypeFunc* ShenandoahBarrierSetC2::clone_barrier_Type() {
532 assert(ShenandoahBarrierSetC2::_clone_barrier_Type != nullptr, "should be initialized");
533 return ShenandoahBarrierSetC2::_clone_barrier_Type;
534 }
535
536 const TypeFunc* ShenandoahBarrierSetC2::load_reference_barrier_Type() {
537 assert(ShenandoahBarrierSetC2::_load_reference_barrier_Type != nullptr, "should be initialized");
538 return ShenandoahBarrierSetC2::_load_reference_barrier_Type;
539 }
540
541 void ShenandoahBarrierSetC2::init() {
542 ShenandoahBarrierSetC2::make_write_barrier_pre_Type();
543 ShenandoahBarrierSetC2::make_clone_barrier_Type();
544 ShenandoahBarrierSetC2::make_load_reference_barrier_Type();
545 }
546
547 void ShenandoahBarrierSetC2::make_write_barrier_pre_Type() {
548 assert(ShenandoahBarrierSetC2::_write_barrier_pre_Type == nullptr, "should be");
549 const Type **fields = TypeTuple::fields(1);
550 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
551 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
552
553 // create result type (range)
554 fields = TypeTuple::fields(0);
555 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
556
557 ShenandoahBarrierSetC2::_write_barrier_pre_Type = TypeFunc::make(domain, range);
558 }
559
560 void ShenandoahBarrierSetC2::make_clone_barrier_Type() {
561 assert(ShenandoahBarrierSetC2::_clone_barrier_Type == nullptr, "should be");
562 const Type **fields = TypeTuple::fields(1);
563 fields[TypeFunc::Parms+0] = TypeOopPtr::NOTNULL; // src oop
564 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
565
566 // create result type (range)
567 fields = TypeTuple::fields(0);
568 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
569
570 ShenandoahBarrierSetC2::_clone_barrier_Type = TypeFunc::make(domain, range);
571 }
572
573 void ShenandoahBarrierSetC2::make_load_reference_barrier_Type() {
574 assert(ShenandoahBarrierSetC2::_load_reference_barrier_Type == nullptr, "should be");
575 const Type **fields = TypeTuple::fields(2);
576 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM; // original field value
577 fields[TypeFunc::Parms+1] = TypeRawPtr::BOTTOM; // original load address
578
579 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
580
581 // create result type (range)
582 fields = TypeTuple::fields(1);
583 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM;
584 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
585
586 ShenandoahBarrierSetC2::_load_reference_barrier_Type = TypeFunc::make(domain, range);
587 }
588
589 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
590 DecoratorSet decorators = access.decorators();
591
592 const TypePtr* adr_type = access.addr().type();
593 Node* adr = access.addr().node();
594
595 bool no_keepalive = (decorators & AS_NO_KEEPALIVE) != 0;
596
597 if (!access.is_oop()) {
598 return BarrierSetC2::store_at_resolved(access, val);
599 }
600
601 if (no_keepalive) {
602 // No keep-alive means no need for the pre-barrier.
603 return BarrierSetC2::store_at_resolved(access, val);
604 }
605
606 if (access.is_parse_access()) {
607 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
608 GraphKit* kit = parse_access.kit();
609
610 uint adr_idx = kit->C->get_alias_index(adr_type);
611 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
612 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
613 static_cast<const TypeOopPtr*>(val.type()), nullptr /* pre_val */, access.type());
614
615 Node* result = BarrierSetC2::store_at_resolved(access, val);
616
617 if (ShenandoahCardBarrier) {
618 const bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
619 const bool is_array = (decorators & IS_ARRAY) != 0;
620 const bool use_precise = is_array || anonymous;
621 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
622 adr, adr_idx, val.node(), access.type(), use_precise);
623 }
624 return result;
625 } else {
626 assert(access.is_opt_access(), "only for optimization passes");
627 assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code");
628 return BarrierSetC2::store_at_resolved(access, val);
629 }
630 }
631
632 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
633 // 1: non-reference load, no additional barrier is needed
634 if (!access.is_oop()) {
635 return BarrierSetC2::load_at_resolved(access, val_type);
636 }
637
638 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
639 DecoratorSet decorators = access.decorators();
640 BasicType type = access.type();
641
642 // 2: apply LRB if needed
643 if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
644 load = new ShenandoahLoadReferenceBarrierNode(nullptr, load, decorators);
645 if (access.is_parse_access()) {
646 load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load);
647 } else {
648 load = static_cast<C2OptAccess &>(access).gvn().transform(load);
649 }
650 }
651
652 // 3: apply keep-alive barrier for java.lang.ref.Reference if needed
653 if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
654 Node* top = Compile::current()->top();
655 Node* adr = access.addr().node();
656 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
657 Node* obj = access.base();
658
659 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
660 bool on_weak_ref = (decorators & (ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF)) != 0;
661 bool keep_alive = (decorators & AS_NO_KEEPALIVE) == 0;
662
663 // If we are reading the value of the referent field of a Reference
664 // object (either by using Unsafe directly or through reflection)
665 // then, if SATB is enabled, we need to record the referent in an
666 // SATB log buffer using the pre-barrier mechanism.
667 // Also we need to add memory barrier to prevent commoning reads
668 // from this field across safepoint since GC can change its value.
669 if (!on_weak_ref || (unknown && (offset == top || obj == top)) || !keep_alive) {
670 return load;
671 }
672
673 assert(access.is_parse_access(), "entry not supported at optimization time");
674 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
675 GraphKit* kit = parse_access.kit();
676 bool mismatched = (decorators & C2_MISMATCHED) != 0;
677 bool is_unordered = (decorators & MO_UNORDERED) != 0;
678 bool in_native = (decorators & IN_NATIVE) != 0;
679 bool need_cpu_mem_bar = !is_unordered || mismatched || in_native;
680
681 if (on_weak_ref) {
682 // Use the pre-barrier to record the value in the referent field
683 satb_write_barrier_pre(kit, false /* do_load */,
684 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */,
685 load /* pre_val */, T_OBJECT);
686 // Add memory barrier to prevent commoning reads from this field
687 // across safepoint since GC can change its value.
688 kit->insert_mem_bar(Op_MemBarCPUOrder);
689 } else if (unknown) {
690 // We do not require a mem bar inside pre_barrier if need_mem_bar
691 // is set: the barriers would be emitted by us.
692 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
693 }
694 }
695
696 return load;
697 }
698
699 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
700 Node* new_val, const Type* value_type) const {
701 GraphKit* kit = access.kit();
702 if (access.is_oop()) {
703 shenandoah_write_barrier_pre(kit, false /* do_load */,
704 nullptr, nullptr, max_juint, nullptr, nullptr,
705 expected_val /* pre_val */, T_OBJECT);
706
707 MemNode::MemOrd mo = access.mem_node_mo();
708 Node* mem = access.memory();
709 Node* adr = access.addr().node();
710 const TypePtr* adr_type = access.addr().type();
711 Node* load_store = nullptr;
712
713 #ifdef _LP64
714 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
715 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
716 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
717 if (ShenandoahCASBarrier) {
718 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
719 } else {
720 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
721 }
722 } else
723 #endif
724 {
725 if (ShenandoahCASBarrier) {
726 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
727 } else {
728 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
729 }
730 }
731
732 access.set_raw_access(load_store);
733 pin_atomic_op(access);
734
735 #ifdef _LP64
736 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
737 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
738 }
739 #endif
740 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, load_store, access.decorators()));
741 if (ShenandoahCardBarrier) {
742 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
743 access.addr().node(), access.alias_idx(), new_val, T_OBJECT, true);
744 }
745 return load_store;
746 }
747 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
748 }
749
750 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
751 Node* new_val, const Type* value_type) const {
752 GraphKit* kit = access.kit();
753 if (access.is_oop()) {
754 shenandoah_write_barrier_pre(kit, false /* do_load */,
755 nullptr, nullptr, max_juint, nullptr, nullptr,
756 expected_val /* pre_val */, T_OBJECT);
757 DecoratorSet decorators = access.decorators();
758 MemNode::MemOrd mo = access.mem_node_mo();
759 Node* mem = access.memory();
760 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
761 Node* load_store = nullptr;
762 Node* adr = access.addr().node();
763 #ifdef _LP64
764 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
765 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
766 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
767 if (ShenandoahCASBarrier) {
768 if (is_weak_cas) {
769 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
770 } else {
771 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
772 }
773 } else {
774 if (is_weak_cas) {
775 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
776 } else {
777 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
778 }
779 }
780 } else
781 #endif
782 {
783 if (ShenandoahCASBarrier) {
784 if (is_weak_cas) {
785 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
786 } else {
787 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
788 }
789 } else {
790 if (is_weak_cas) {
791 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
792 } else {
793 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
794 }
795 }
796 }
797 access.set_raw_access(load_store);
798 pin_atomic_op(access);
799 if (ShenandoahCardBarrier) {
800 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
801 access.addr().node(), access.alias_idx(), new_val, T_OBJECT, true);
802 }
803 return load_store;
804 }
805 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
806 }
807
808 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
809 GraphKit* kit = access.kit();
810 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
811 if (access.is_oop()) {
812 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, result, access.decorators()));
813 shenandoah_write_barrier_pre(kit, false /* do_load */,
814 nullptr, nullptr, max_juint, nullptr, nullptr,
815 result /* pre_val */, T_OBJECT);
816 if (ShenandoahCardBarrier) {
817 post_barrier(kit, kit->control(), access.raw_access(), access.base(),
818 access.addr().node(), access.alias_idx(), val, T_OBJECT, true);
819 }
820 }
821 return result;
822 }
823
824
825 bool ShenandoahBarrierSetC2::is_gc_pre_barrier_node(Node* node) const {
826 return is_shenandoah_wb_pre_call(node);
827 }
828
829 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
830 return (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) ||
831 is_shenandoah_lrb_call(node) ||
832 is_shenandoah_wb_pre_call(node) ||
833 is_shenandoah_clone_call(node);
834 }
835
836 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
837 if (c == nullptr) {
838 return c;
839 }
840 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
841 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
842 }
843 return c;
844 }
845
846 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
847 return !ShenandoahBarrierC2Support::expand(C, igvn);
848 }
849
850 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
851 if (mode == LoopOptsShenandoahExpand) {
852 assert(UseShenandoahGC, "only for shenandoah");
853 ShenandoahBarrierC2Support::pin_and_expand(phase);
854 return true;
855 }
856 return false;
857 }
858
859 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, bool is_clone_instance, ArrayCopyPhase phase) const {
860 bool is_oop = is_reference_type(type);
861 if (!is_oop) {
862 return false;
863 }
864 if (ShenandoahSATBBarrier && tightly_coupled_alloc) {
865 if (phase == Optimization) {
866 return false;
867 }
868 return !is_clone;
869 }
870 return true;
871 }
872
873 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) {
874 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr();
875 if (src_type->isa_instptr() != nullptr) {
876 ciInstanceKlass* ik = src_type->is_instptr()->instance_klass();
877 if ((src_type->klass_is_exact() || !ik->has_subklass()) && !ik->has_injected_fields()) {
878 if (ik->has_object_fields()) {
879 return true;
880 } else {
881 if (!src_type->klass_is_exact()) {
882 Compile::current()->dependencies()->assert_leaf_type(ik);
883 }
884 }
885 } else {
886 return true;
887 }
888 } else if (src_type->isa_aryptr()) {
889 BasicType src_elem = src_type->isa_aryptr()->elem()->array_element_basic_type();
890 if (is_reference_type(src_elem, true)) {
891 return true;
892 }
893 } else {
894 return true;
895 }
896 return false;
897 }
898
899 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
900 Node* ctrl = ac->in(TypeFunc::Control);
901 Node* mem = ac->in(TypeFunc::Memory);
902 Node* src_base = ac->in(ArrayCopyNode::Src);
903 Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
904 Node* dest_base = ac->in(ArrayCopyNode::Dest);
905 Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
906 Node* length = ac->in(ArrayCopyNode::Length);
907
908 Node* src = phase->basic_plus_adr(src_base, src_offset);
909 Node* dest = phase->basic_plus_adr(dest_base, dest_offset);
910
911 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) {
912 // Check if heap is has forwarded objects. If it does, we need to call into the special
913 // routine that would fix up source references before we can continue.
914
915 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
916 Node* region = new RegionNode(PATH_LIMIT);
917 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM);
918
919 Node* thread = phase->transform_later(new ThreadLocalNode());
920 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
921 Node* gc_state_addr = phase->transform_later(AddPNode::make_off_heap(thread, offset));
922
923 uint gc_state_idx = Compile::AliasIdxRaw;
924 const TypePtr* gc_state_adr_type = nullptr; // debug-mode-only argument
925 DEBUG_ONLY(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
926
927 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered));
928 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED)));
929 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT)));
930 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne));
931
932 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If();
933 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff));
934 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff));
935
936 // Heap is stable, no need to do anything additional
937 region->init_req(_heap_stable, stable_ctrl);
938 mem_phi->init_req(_heap_stable, mem);
939
940 // Heap is unstable, call into clone barrier stub
941 Node* call = phase->make_leaf_call(unstable_ctrl, mem,
942 ShenandoahBarrierSetC2::clone_barrier_Type(),
943 CAST_FROM_FN_PTR(address, ShenandoahRuntime::clone_barrier),
944 "shenandoah_clone",
945 TypeRawPtr::BOTTOM,
946 src_base);
947 call = phase->transform_later(call);
948
949 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control));
950 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory));
951 region->init_req(_heap_unstable, ctrl);
952 mem_phi->init_req(_heap_unstable, mem);
953
954 // Wire up the actual arraycopy stub now
955 ctrl = phase->transform_later(region);
956 mem = phase->transform_later(mem_phi);
957
958 const char* name = "arraycopy";
959 call = phase->make_leaf_call(ctrl, mem,
960 OptoRuntime::fast_arraycopy_Type(),
961 phase->basictype2arraycopy(T_LONG, nullptr, nullptr, true, name, true),
962 name, TypeRawPtr::BOTTOM,
963 src, dest, length
964 LP64_ONLY(COMMA phase->top()));
965 call = phase->transform_later(call);
966
967 // Hook up the whole thing into the graph
968 phase->igvn().replace_node(ac, call);
969 } else {
970 BarrierSetC2::clone_at_expansion(phase, ac);
971 }
972 }
973
974
975 // Support for macro expanded GC barriers
976 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
977 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
978 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
979 }
980 }
981
982 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
983 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
984 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node);
985 }
986 }
987
988 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const {
989 if (is_shenandoah_wb_pre_call(node)) {
990 shenandoah_eliminate_wb_pre(node, ¯o->igvn());
991 }
992 if (ShenandoahCardBarrier && node->Opcode() == Op_CastP2X) {
993 for (DUIterator_Last imin, i = node->last_outs(imin); i >= imin; --i) {
994 Node* shift = node->last_out(i);
995 for (DUIterator_Last kmin, k = shift->last_outs(kmin); k >= kmin; --k) {
996 Node* addp = shift->last_out(k);
997 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
998 Node* mem = addp->last_out(j);
999 if (UseCondCardMark && mem->is_Load()) {
1000 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
1001 // The load is checking if the card has been written so
1002 // replace it with zero to fold the test.
1003 macro->replace_node(mem, macro->intcon(0));
1004 continue;
1005 }
1006 assert(mem->is_Store(), "store required");
1007 macro->replace_node(mem, mem->in(MemNode::Memory));
1008 }
1009 }
1010 }
1011 }
1012 }
1013
1014 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
1015 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
1016 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
1017 c = c->unique_ctrl_out();
1018 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
1019 c = c->unique_ctrl_out();
1020 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
1021 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
1022 assert(iff->is_If(), "expect test");
1023 if (!is_shenandoah_marking_if(igvn, iff)) {
1024 c = c->unique_ctrl_out();
1025 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
1026 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
1027 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
1028 }
1029 Node* cmpx = iff->in(1)->in(1);
1030 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
1031 igvn->rehash_node_delayed(call);
1032 call->del_req(call->req()-1);
1033 }
1034
1035 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
1036 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
1037 igvn->add_users_to_worklist(node);
1038 }
1039 }
1040
1041 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
1042 for (uint i = 0; i < useful.size(); i++) {
1043 Node* n = useful.at(i);
1044 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
1045 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1046 C->record_for_igvn(n->fast_out(i));
1047 }
1048 }
1049 }
1050
1051 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) {
1052 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i);
1053 if (!useful.member(n)) {
1054 state()->remove_load_reference_barrier(n);
1055 }
1056 }
1057 }
1058
1059 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
1060 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
1061 }
1062
1063 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
1064 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
1065 }
1066
1067 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
1068 // expanded later, then now is the time to do so.
1069 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
1070
1071 #ifdef ASSERT
1072 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
1073 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) {
1074 ShenandoahBarrierC2Support::verify(Compile::current()->root());
1075 } else if (phase == BarrierSetC2::BeforeCodeGen) {
1076 // Verify Shenandoah pre-barriers
1077 const int gc_state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
1078
1079 Unique_Node_List visited;
1080 Node_List worklist;
1081 // We're going to walk control flow backwards starting from the Root
1082 worklist.push(compile->root());
1083 while (worklist.size() > 0) {
1084 Node *x = worklist.pop();
1085 if (x == nullptr || x == compile->top()) {
1086 continue;
1087 }
1088
1089 if (visited.member(x)) {
1090 continue;
1091 } else {
1092 visited.push(x);
1093 }
1094
1095 if (x->is_Region()) {
1096 for (uint i = 1; i < x->req(); i++) {
1097 worklist.push(x->in(i));
1098 }
1099 } else {
1100 worklist.push(x->in(0));
1101 // We are looking for the pattern:
1102 // /->ThreadLocal
1103 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
1104 // \->ConI(0)
1105 // We want to verify that the If and the LoadB have the same control
1106 // See GraphKit::g1_write_barrier_pre()
1107 if (x->is_If()) {
1108 IfNode *iff = x->as_If();
1109 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
1110 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
1111 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
1112 && cmp->in(1)->is_Load()) {
1113 LoadNode *load = cmp->in(1)->as_Load();
1114 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
1115 && load->in(2)->in(3)->is_Con()
1116 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == gc_state_offset) {
1117
1118 Node *if_ctrl = iff->in(0);
1119 Node *load_ctrl = load->in(0);
1120
1121 if (if_ctrl != load_ctrl) {
1122 // Skip possible CProj->NeverBranch in infinite loops
1123 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
1124 && if_ctrl->in(0)->is_NeverBranch()) {
1125 if_ctrl = if_ctrl->in(0)->in(0);
1126 }
1127 }
1128 assert(load_ctrl != nullptr && if_ctrl == load_ctrl, "controls must match");
1129 }
1130 }
1131 }
1132 }
1133 }
1134 }
1135 }
1136 }
1137 #endif
1138
1139 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
1140 if (is_shenandoah_wb_pre_call(n)) {
1141 uint cnt = ShenandoahBarrierSetC2::write_barrier_pre_Type()->domain()->cnt();
1142 if (n->req() > cnt) {
1143 Node* addp = n->in(cnt);
1144 if (has_only_shenandoah_wb_pre_uses(addp)) {
1145 n->del_req(cnt);
1146 if (can_reshape) {
1147 phase->is_IterGVN()->_worklist.push(addp);
1148 }
1149 return n;
1150 }
1151 }
1152 }
1153 if (n->Opcode() == Op_CmpP) {
1154 Node* in1 = n->in(1);
1155 Node* in2 = n->in(2);
1156
1157 // If one input is null, then step over the strong LRB barriers on the other input
1158 if (in1->bottom_type() == TypePtr::NULL_PTR &&
1159 !((in2->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1160 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in2)->decorators()))) {
1161 in2 = step_over_gc_barrier(in2);
1162 }
1163 if (in2->bottom_type() == TypePtr::NULL_PTR &&
1164 !((in1->Opcode() == Op_ShenandoahLoadReferenceBarrier) &&
1165 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in1)->decorators()))) {
1166 in1 = step_over_gc_barrier(in1);
1167 }
1168
1169 if (in1 != n->in(1)) {
1170 n->set_req_X(1, in1, phase);
1171 assert(in2 == n->in(2), "only one change");
1172 return n;
1173 }
1174 if (in2 != n->in(2)) {
1175 n->set_req_X(2, in2, phase);
1176 return n;
1177 }
1178 } else if (can_reshape &&
1179 n->Opcode() == Op_If &&
1180 ShenandoahBarrierC2Support::is_heap_stable_test(n) &&
1181 n->in(0) != nullptr &&
1182 n->outcnt() == 2) {
1183 Node* dom = n->in(0);
1184 Node* prev_dom = n;
1185 int op = n->Opcode();
1186 int dist = 16;
1187 // Search up the dominator tree for another heap stable test
1188 while (dom->Opcode() != op || // Not same opcode?
1189 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1?
1190 prev_dom->in(0) != dom) { // One path of test does not dominate?
1191 if (dist < 0) return nullptr;
1192
1193 dist--;
1194 prev_dom = dom;
1195 dom = IfNode::up_one_dom(dom);
1196 if (!dom) return nullptr;
1197 }
1198
1199 // Check that we did not follow a loop back to ourselves
1200 if (n == dom) {
1201 return nullptr;
1202 }
1203
1204 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN(), false);
1205 }
1206
1207 return nullptr;
1208 }
1209
1210 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1211 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1212 Node* u = n->fast_out(i);
1213 if (!is_shenandoah_wb_pre_call(u)) {
1214 return false;
1215 }
1216 }
1217 return n->outcnt() > 0;
1218 }
1219
1220 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode, Unique_Node_List& dead_nodes) const {
1221 switch (opcode) {
1222 case Op_CallLeaf:
1223 case Op_CallLeafNoFP: {
1224 assert (n->is_Call(), "");
1225 CallNode *call = n->as_Call();
1226 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1227 uint cnt = ShenandoahBarrierSetC2::write_barrier_pre_Type()->domain()->cnt();
1228 if (call->req() > cnt) {
1229 assert(call->req() == cnt + 1, "only one extra input");
1230 Node *addp = call->in(cnt);
1231 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1232 call->del_req(cnt);
1233 }
1234 }
1235 return false;
1236 }
1237 case Op_ShenandoahCompareAndSwapP:
1238 case Op_ShenandoahCompareAndSwapN:
1239 case Op_ShenandoahWeakCompareAndSwapN:
1240 case Op_ShenandoahWeakCompareAndSwapP:
1241 case Op_ShenandoahCompareAndExchangeP:
1242 case Op_ShenandoahCompareAndExchangeN:
1243 return true;
1244 case Op_ShenandoahLoadReferenceBarrier:
1245 assert(false, "should have been expanded already");
1246 return true;
1247 default:
1248 return false;
1249 }
1250 }
1251
1252 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1253 switch (opcode) {
1254 case Op_ShenandoahCompareAndExchangeP:
1255 case Op_ShenandoahCompareAndExchangeN:
1256 conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1257 // fallthrough
1258 case Op_ShenandoahWeakCompareAndSwapP:
1259 case Op_ShenandoahWeakCompareAndSwapN:
1260 case Op_ShenandoahCompareAndSwapP:
1261 case Op_ShenandoahCompareAndSwapN:
1262 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1263 return true;
1264 case Op_StoreP: {
1265 Node* adr = n->in(MemNode::Address);
1266 const Type* adr_type = gvn->type(adr);
1267 // Pointer stores in Shenandoah barriers looks like unsafe access.
1268 // Ignore such stores to be able scalar replace non-escaping
1269 // allocations.
1270 if (adr_type->isa_rawptr() && adr->is_AddP()) {
1271 Node* base = conn_graph->get_addp_base(adr);
1272 if (base->Opcode() == Op_LoadP &&
1273 base->in(MemNode::Address)->is_AddP()) {
1274 adr = base->in(MemNode::Address);
1275 Node* tls = conn_graph->get_addp_base(adr);
1276 if (tls->Opcode() == Op_ThreadLocal) {
1277 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1278 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1279 if (offs == buf_offset) {
1280 return true; // Pre barrier previous oop value store.
1281 }
1282 }
1283 }
1284 }
1285 return false;
1286 }
1287 case Op_ShenandoahLoadReferenceBarrier:
1288 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist);
1289 return true;
1290 default:
1291 // Nothing
1292 break;
1293 }
1294 return false;
1295 }
1296
1297 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1298 switch (opcode) {
1299 case Op_ShenandoahCompareAndExchangeP:
1300 case Op_ShenandoahCompareAndExchangeN: {
1301 Node *adr = n->in(MemNode::Address);
1302 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, nullptr);
1303 // fallthrough
1304 }
1305 case Op_ShenandoahCompareAndSwapP:
1306 case Op_ShenandoahCompareAndSwapN:
1307 case Op_ShenandoahWeakCompareAndSwapP:
1308 case Op_ShenandoahWeakCompareAndSwapN:
1309 return conn_graph->add_final_edges_unsafe_access(n, opcode);
1310 case Op_ShenandoahLoadReferenceBarrier:
1311 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), nullptr);
1312 return true;
1313 default:
1314 // Nothing
1315 break;
1316 }
1317 return false;
1318 }
1319
1320 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1321 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1322 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1323
1324 }
1325
1326 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1327 switch (opcode) {
1328 case Op_ShenandoahCompareAndExchangeP:
1329 case Op_ShenandoahCompareAndExchangeN:
1330 case Op_ShenandoahWeakCompareAndSwapP:
1331 case Op_ShenandoahWeakCompareAndSwapN:
1332 case Op_ShenandoahCompareAndSwapP:
1333 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree
1334 Node* newval = n->in(MemNode::ValueIn);
1335 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1336 Node* pair = new BinaryNode(oldval, newval);
1337 n->set_req(MemNode::ValueIn,pair);
1338 n->del_req(LoadStoreConditionalNode::ExpectedIn);
1339 return true;
1340 }
1341 default:
1342 break;
1343 }
1344 return false;
1345 }
1346
1347 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1348 return xop == Op_ShenandoahCompareAndExchangeP ||
1349 xop == Op_ShenandoahCompareAndExchangeN ||
1350 xop == Op_ShenandoahWeakCompareAndSwapP ||
1351 xop == Op_ShenandoahWeakCompareAndSwapN ||
1352 xop == Op_ShenandoahCompareAndSwapN ||
1353 xop == Op_ShenandoahCompareAndSwapP;
1354 }
|
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "classfile/javaClasses.inline.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
29 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
30 #include "gc/shenandoah/shenandoahForwarding.hpp"
31 #include "gc/shenandoah/shenandoahHeap.hpp"
32 #include "gc/shenandoah/shenandoahRuntime.hpp"
33 #include "gc/shenandoah/shenandoahThreadLocalData.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/narrowptrnode.hpp"
40 #include "opto/output.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 BarrierSetC2State(comp_arena),
50 _stubs(new (comp_arena) GrowableArray<ShenandoahBarrierStubC2*>(comp_arena, 8, 0, nullptr)),
51 _trampoline_stubs_count(0),
52 _stubs_start_offset(0),
53 _stubs_current_total_size(0) {
54 }
55
56 static void set_barrier_data(C2Access& access, bool load, bool store) {
57 if (!access.is_oop()) {
58 return;
59 }
60
61 DecoratorSet decorators = access.decorators();
62 bool tightly_coupled = (decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0;
63 bool in_heap = (decorators & IN_HEAP) != 0;
64 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
65 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
66
67 if (tightly_coupled) {
68 access.set_barrier_data(ShenandoahBitElided);
69 return;
70 }
71
72 uint8_t barrier_data = 0;
73
74 if (load) {
75 if (ShenandoahLoadRefBarrier) {
76 if (on_phantom) {
77 barrier_data |= ShenandoahBitPhantom;
78 } else if (on_weak) {
79 barrier_data |= ShenandoahBitWeak;
80 } else {
81 barrier_data |= ShenandoahBitStrong;
82 }
83 }
84 }
85
86 if (store) {
87 if (ShenandoahSATBBarrier) {
88 barrier_data |= ShenandoahBitKeepAlive;
89 }
90 if (ShenandoahCardBarrier && in_heap) {
91 barrier_data |= ShenandoahBitCardMark;
92 }
93 }
94
95 if (!in_heap) {
96 barrier_data |= ShenandoahBitNative;
97 }
98
99 access.set_barrier_data(barrier_data);
100 }
101
102 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
103 // 1: Non-reference load, no additional barrier is needed
104 if (!access.is_oop()) {
105 return BarrierSetC2::load_at_resolved(access, val_type);
106 }
107
108 // 2. Set barrier data for load
109 set_barrier_data(access, /* load = */ true, /* store = */ false);
110
111 // 3. Correction: If we are reading the value of the referent field of
112 // a Reference object, we need to record the referent resurrection.
113 DecoratorSet decorators = access.decorators();
114 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
115 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
116 bool no_keepalive = (decorators & AS_NO_KEEPALIVE) != 0;
117 bool needs_keepalive = ((on_weak || on_phantom) && !no_keepalive);
118 if (needs_keepalive) {
119 uint8_t barriers = access.barrier_data() | (ShenandoahSATBBarrier ? ShenandoahBitKeepAlive : 0);
120 access.set_barrier_data(barriers);
121 }
122
123 return BarrierSetC2::load_at_resolved(access, val_type);
124 }
125
126 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
127 // 1: Non-reference store, no additional barrier is needed
128 if (!access.is_oop()) {
129 return BarrierSetC2::store_at_resolved(access, val);
130 }
131
132 // 2. Set barrier data for store
133 set_barrier_data(access, /* load = */ false, /* store = */ true);
134
135 // 3. Correction: avoid keep-alive barriers that should not do keep-alive.
136 DecoratorSet decorators = access.decorators();
137 bool no_keepalive = (decorators & AS_NO_KEEPALIVE) != 0;
138 if (no_keepalive) {
139 access.set_barrier_data(access.barrier_data() & ~ShenandoahBitKeepAlive);
140 }
141
142 return BarrierSetC2::store_at_resolved(access, val);
143 }
144
145 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
146 Node* new_val, const Type* value_type) const {
147 set_barrier_data(access, /* load = */ true, /* store = */ true);
148 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
149 }
150
151 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
152 Node* new_val, const Type* value_type) const {
153 set_barrier_data(access, /* load = */ true, /* store = */ true);
154 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
155 }
156
157 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
158 set_barrier_data(access, /* load = */ true, /* store = */ true);
159 return BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
160 }
161
162 bool ShenandoahBarrierSetC2::is_Load(int opcode) {
163 switch (opcode) {
164 case Op_LoadN:
165 case Op_LoadP:
166 return true;
167 default:
168 return false;
169 }
170 }
171
172 bool ShenandoahBarrierSetC2::is_Store(int opcode) {
173 switch (opcode) {
174 case Op_StoreN:
175 case Op_StoreP:
176 return true;
177 default:
178 return false;
179 }
180 }
181
182 bool ShenandoahBarrierSetC2::is_LoadStore(int opcode) {
183 switch (opcode) {
184 case Op_CompareAndExchangeN:
185 case Op_CompareAndExchangeP:
186 case Op_WeakCompareAndSwapN:
187 case Op_WeakCompareAndSwapP:
188 case Op_CompareAndSwapN:
189 case Op_CompareAndSwapP:
190 case Op_GetAndSetP:
191 case Op_GetAndSetN:
192 return true;
193 default:
194 return false;
195 }
196 }
197
198 bool ShenandoahBarrierSetC2::can_remove_load_barrier(Node* root) {
199 // Check if all outs feed into nodes that do not expose the oops to the rest
200 // of the runtime system. In this case, we can elide the LRB barrier. We bail
201 // out with false at the first sight of trouble.
202
203 ResourceMark rm;
204 VectorSet visited;
205 Node_List worklist;
206 worklist.push(root);
207
208 while (worklist.size() > 0) {
209 Node* n = worklist.pop();
210 if (visited.test_set(n->_idx)) {
211 continue;
212 }
213
214 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
215 Node* out = n->fast_out(i);
216 switch (out->Opcode()) {
217 case Op_Phi:
218 case Op_EncodeP:
219 case Op_DecodeN:
220 case Op_CastPP:
221 case Op_CheckCastPP:
222 case Op_AddP: {
223 // Transitive node, check if any other outs are doing anything troublesome.
224 worklist.push(out);
225 break;
226 }
227
228 case Op_LoadRange: {
229 // Array length is the same in all copies.
230 break;
231 }
232
233 case Op_LoadKlass: {
234 // Klass is the same in all copies.
235 // We would have liked to assert -UCOH, but there are legitimate klass
236 // loads from native Klass* instances, which are also safe under +UCOH.
237 break;
238 }
239
240 case Op_LoadNKlass: {
241 // Similar to above, but LoadNKlass is only safe without +UCOH.
242 // With +UCOH, it loads from mark word, which clashes with forwarding pointers.
243 if (!UseCompactObjectHeaders) {
244 break;
245 }
246 return false;
247 }
248
249 case Op_CmpN: {
250 if (out->in(1) == n &&
251 out->in(2)->Opcode() == Op_ConN &&
252 out->in(2)->get_narrowcon() == 0) {
253 // Null check, no oop is exposed.
254 break;
255 }
256 if (out->in(2) == n &&
257 out->in(1)->Opcode() == Op_ConN &&
258 out->in(1)->get_narrowcon() == 0) {
259 // Null check, no oop is exposed.
260 break;
261 }
262 return false;
263 }
264
265 case Op_CmpP: {
266 if (out->in(1) == n &&
267 out->in(2)->Opcode() == Op_ConP &&
268 out->in(2)->get_ptr() == 0) {
269 // Null check, no oop is exposed.
270 break;
271 }
272 if (out->in(2) == n &&
273 out->in(1)->Opcode() == Op_ConP &&
274 out->in(1)->get_ptr() == 0) {
275 // Null check, no oop is exposed.
276 break;
277 }
278 return false;
279 }
280
281 case Op_CallStaticJava: {
282 if (out->as_CallStaticJava()->is_uncommon_trap()) {
283 // Local feeds into uncommon trap. Deopt machinery handles barriers itself.
284 break;
285 }
286 return false;
287 }
288
289 default: {
290 // Paranoidly distrust any other nodes.
291 return false;
292 }
293 }
294 }
295 }
296
297 // Nothing troublesome found.
298 return true;
299 }
300
301 uint8_t ShenandoahBarrierSetC2::refine_load(Node* n, uint8_t bd) {
302 assert(ShenandoahElideIdealBarriers, "Checked by caller");
303 assert(bd != 0, "Checked by caller");
304
305 // Do not touch weak loads at all: they are responsible for shielding from
306 // Reference.referent resurrection.
307 if ((bd & (ShenandoahBitWeak | ShenandoahBitPhantom)) != 0) {
308 return bd;
309 }
310
311 if (((bd & ShenandoahBitStrong) != 0) && can_remove_load_barrier(n)) {
312 bd &= ~ShenandoahBitStrong;
313 }
314
315 return bd;
316 }
317
318 uint8_t ShenandoahBarrierSetC2::refine_store(Node* n, uint8_t bd) {
319 assert(ShenandoahElideIdealBarriers, "Checked by caller");
320 assert(bd != 0, "Checked by caller");
321 assert(n->is_Mem() || n->is_LoadStore(), "Sanity");
322
323 const Node* newval = n->in(MemNode::ValueIn);
324 assert(newval != nullptr, "Should be present");
325
326 // Type system tells us something about nullity?
327 const Type* newval_bottom = newval->bottom_type();
328 assert(newval_bottom->isa_oopptr() || newval_bottom->isa_narrowoop() ||
329 newval_bottom == TypePtr::NULL_PTR, "Should be an oop store");
330 const TypePtr* newval_type = newval_bottom->make_ptr();
331 assert(newval_type != nullptr, "Should have been filtered before");
332 TypePtr::PTR newval_type_ptr = newval_type->ptr();
333 if (newval_type_ptr == TypePtr::Null) {
334 bd &= ~ShenandoahBitNotNull;
335 // Card table barrier is not needed if we store null.
336 bd &= ~ShenandoahBitCardMark;
337 } else if (newval_type_ptr == TypePtr::NotNull) {
338 // Definitely not null.
339 bd |= ShenandoahBitNotNull;
340 }
341
342 return bd;
343 }
344
345 void ShenandoahBarrierSetC2::final_refinement(Compile* compile) const {
346 ResourceMark rm;
347 Unique_Node_List wq;
348
349 RootNode* root = compile->root();
350 wq.push(root);
351
352 // Also seed the outs to capture nodes are not reachable from in()-s, e.g. endless loops.
353 for (DUIterator_Fast imax, i = root->fast_outs(imax); i < imax; i++) {
354 Node* m = root->fast_out(i);
355 wq.push(m);
356 }
357
358 for (uint next = 0; next < wq.size(); next++) {
359 Node* n = wq.at(next);
360
361 assert(!n->is_Mach(), "No Mach nodes here yet");
362
363 int opc = n->Opcode();
364 bool is_load = is_Load(opc);
365 bool is_store = is_Store(opc);
366 bool is_load_store = is_LoadStore(opc);
367
368 uint8_t orig_bd = 0;
369 if (is_load_store) {
370 orig_bd = n->as_LoadStore()->barrier_data();
371 } else if (is_load || is_store) {
372 orig_bd = n->as_Mem()->barrier_data();
373 }
374
375 uint8_t bd = orig_bd;
376 if (ShenandoahElideIdealBarriers && bd != 0) {
377 // Note: we cannot apply load optimizations to LoadStores,
378 // because their load barriers are needed for fixups.
379 if (is_load) {
380 bd = refine_load(n, bd);
381 }
382 if (is_store || is_load_store) {
383 bd = refine_store(n, bd);
384 }
385 }
386
387 // If there are no real barrier flags on the node, strip away additional fluff.
388 // Matcher does not care about this, and we would like to avoid invoking "barrier_data() != 0"
389 // rules when the only flags are the irrelevant fluff.
390 if ((bd != 0) && (bd & ShenandoahBitsReal) == 0) {
391 bd = 0;
392 }
393
394 if (bd != orig_bd) {
395 if (is_load_store) {
396 n->as_LoadStore()->set_barrier_data(bd);
397 } else {
398 n->as_Mem()->set_barrier_data(bd);
399 }
400 }
401
402 for (uint j = 0; j < n->req(); j++) {
403 Node* in = n->in(j);
404 if (in != nullptr) {
405 wq.push(in);
406 }
407 }
408 }
409 }
410
411 // Support for macro expanded GC barriers
412 void ShenandoahBarrierSetC2::eliminate_gc_barrier_data(Node* node) const {
413 if (node->is_LoadStore()) {
414 LoadStoreNode* loadstore = node->as_LoadStore();
415 loadstore->set_barrier_data(0);
416 } else if (node->is_Mem()) {
417 MemNode* mem = node->as_Mem();
418 mem->set_barrier_data(0);
419 }
420 }
421
422 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const {
423 eliminate_gc_barrier_data(node);
424 }
425
426 void ShenandoahBarrierSetC2::elide_dominated_barrier(MachNode* node, MachNode* dominator) const {
427 uint8_t orig_bd = node->barrier_data();
428 if (orig_bd == 0) {
429 // Nothing to do.
430 return;
431 }
432
433 uint8_t bd = orig_bd;
434 int node_opcode = node->ideal_Opcode();
435
436 if (dominator == nullptr) {
437 // Must be allocation node.
438 if (is_Load(node_opcode) || is_LoadStore(node_opcode)) {
439 // Loads from recent allocations do not need LRBs.
440 bd &= ~ShenandoahBitStrong;
441 }
442 if (is_Store(node_opcode) || is_LoadStore(node_opcode)) {
443 // Stores to recent allocations do not need KA or CM.
444 bd &= ~ShenandoahBitKeepAlive;
445 bd &= ~ShenandoahBitCardMark;
446 }
447 } else {
448 // LoadStores do not get these optimizations, since their LRBs
449 // are required for fixups.
450 if (is_Load(node_opcode) || is_Store(node_opcode)) {
451 int dom_opcode = dominator->ideal_Opcode();
452 uint8_t dom_bd = dominator->barrier_data();
453
454 if (is_Load(dom_opcode) || is_LoadStore(dom_opcode)) {
455 // If dominating load is set up to perform LRB fixups, no further LRB is needed.
456 if ((dom_bd & ShenandoahBitStrong) != 0) {
457 bd &= ~ShenandoahBitStrong;
458 }
459 }
460 if (is_Store(dom_opcode)) {
461 // Dominating store has stored the good ref, no LRB is needed.
462 bd &= ~ShenandoahBitStrong;
463 }
464 }
465 }
466
467 if (orig_bd != bd) {
468 // We are already in final output.
469 // Strip the extra barrier data if no real bits are left.
470 if ((bd & ShenandoahBitsReal) != 0) {
471 node->set_barrier_data(bd);
472 } else {
473 node->set_barrier_data(0);
474 }
475 }
476 }
477
478 void ShenandoahBarrierSetC2::analyze_dominating_barriers() const {
479 if (!ShenandoahElideMachBarriers) {
480 return;
481 }
482
483 ResourceMark rm;
484 Node_List accesses, dominators;
485
486 PhaseCFG* const cfg = Compile::current()->cfg();
487 for (uint i = 0; i < cfg->number_of_blocks(); ++i) {
488 const Block* const block = cfg->get_block(i);
489 for (uint j = 0; j < block->number_of_nodes(); ++j) {
490 Node* const node = block->get_node(j);
491
492 // Everything that happens in allocations does not need barriers.
493 // Record them for dominance analysis.
494 if (node->is_Phi() && is_allocation(node)) {
495 dominators.push(node);
496 continue;
497 }
498
499 if (!node->is_Mach()) {
500 continue;
501 }
502
503 MachNode* const mach = node->as_Mach();
504 int opcode = mach->ideal_Opcode();
505 if (is_Load(opcode) || is_Store(opcode) || is_LoadStore(opcode)) {
506 if ((mach->barrier_data() & ShenandoahBitsReal) != 0) {
507 accesses.push(mach);
508 dominators.push(mach);
509 }
510 }
511 }
512 }
513
514 elide_dominated_barriers(accesses, dominators);
515 }
516
517 uint ShenandoahBarrierSetC2::estimated_barrier_size(const Node* node) const {
518 // Barrier impact on fast-path is driven by GC state checks emitted very late.
519 // These checks are tight load-test-branch sequences, with no impact on C2 graph
520 // size. Limiting unrolling in presence of GC barriers might turn some loops
521 // tighter than with default unrolling, which may benefit performance due to denser
522 // code. Testing shows it is still counter-productive.
523 // Therefore, we report zero barrier size to let C2 do its normal thing.
524 return 0;
525 }
526
527 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, bool is_clone_instance, ArrayCopyPhase phase) const {
528 bool is_oop = is_reference_type(type);
529 if (!is_oop) {
530 return false;
531 }
532 if (ShenandoahSATBBarrier && tightly_coupled_alloc) {
533 if (phase == Optimization) {
534 return false;
535 }
536 return !is_clone;
537 }
538 return true;
539 }
540
541 bool ShenandoahBarrierSetC2::clone_needs_barrier(const TypeOopPtr* src_type, bool& is_oop_array) {
542 if (!ShenandoahCloneBarrier) {
543 return false;
544 }
545
546 if (src_type->isa_instptr() != nullptr) {
547 // Instance: need barrier only if there is a possibility of having an oop anywhere in it.
548 ciInstanceKlass* ik = src_type->is_instptr()->instance_klass();
549 if ((src_type->klass_is_exact() || !ik->has_subklass()) &&
550 !ik->has_injected_fields() && !ik->has_object_fields()) {
551 if (!src_type->klass_is_exact()) {
552 // Class is *currently* the leaf in the hierarchy.
553 // Record the dependency so that we deopt if this does not hold in future.
554 Compile::current()->dependencies()->assert_leaf_type(ik);
555 }
556 return false;
557 }
558 } else if (src_type->isa_aryptr() != nullptr) {
559 // Array: need barrier only if array is oop-bearing.
560 BasicType src_elem = src_type->isa_aryptr()->elem()->array_element_basic_type();
561 if (is_reference_type(src_elem, true)) {
562 is_oop_array = true;
563 } else {
564 return false;
565 }
566 }
567
568 // Assume the worst.
569 return true;
570 }
571
572 void ShenandoahBarrierSetC2::clone(GraphKit* kit, Node* src_base, Node* dst_base, Node* size, bool is_array) const {
573 const TypeOopPtr* src_type = kit->gvn().type(src_base)->is_oopptr();
574
575 bool is_oop_array = false;
576 if (!clone_needs_barrier(src_type, is_oop_array)) {
577 // No barrier is needed? Just do what common BarrierSetC2 wants with it.
578 BarrierSetC2::clone(kit, src_base, dst_base, size, is_array);
579 return;
580 }
581
582 if (ShenandoahCloneRuntime || !is_array || !is_oop_array) {
583 // Looks like an instance? Prepare the instance clone. This would either
584 // be exploded into individual accesses or be left as runtime call.
585 // Common BarrierSetC2 prepares everything for both cases.
586 BarrierSetC2::clone(kit, src_base, dst_base, size, is_array);
587 return;
588 }
589
590 // We are cloning the oop array. Prepare to call the normal arraycopy stub
591 // after the expansion. Normal stub takes the number of actual type-sized
592 // elements to copy after the base, compute the count here.
593 Node* offset = kit->MakeConX(arrayOopDesc::base_offset_in_bytes(UseCompressedOops ? T_NARROWOOP : T_OBJECT));
594 size = kit->gvn().transform(new SubXNode(size, offset));
595 size = kit->gvn().transform(new URShiftXNode(size, kit->intcon(LogBytesPerHeapOop)));
596 ArrayCopyNode* ac = ArrayCopyNode::make(kit, false, src_base, offset, dst_base, offset, size, true, false);
597 ac->set_clone_array();
598 Node* n = kit->gvn().transform(ac);
599 if (n == ac) {
600 ac->set_adr_type(TypeRawPtr::BOTTOM);
601 kit->set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), TypeRawPtr::BOTTOM);
602 } else {
603 kit->set_all_memory(n);
604 }
605 }
606
607 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
608 Node* const ctrl = ac->in(TypeFunc::Control);
609 Node* const mem = ac->in(TypeFunc::Memory);
610 Node* const src = ac->in(ArrayCopyNode::Src);
611 Node* const src_offset = ac->in(ArrayCopyNode::SrcPos);
612 Node* const dest = ac->in(ArrayCopyNode::Dest);
613 Node* const dest_offset = ac->in(ArrayCopyNode::DestPos);
614 Node* length = ac->in(ArrayCopyNode::Length);
615
616 const TypeOopPtr* src_type = phase->igvn().type(src)->is_oopptr();
617
618 bool is_oop_array = false;
619 if (!clone_needs_barrier(src_type, is_oop_array)) {
620 // No barrier is needed? Expand to normal HeapWord-sized arraycopy.
621 BarrierSetC2::clone_at_expansion(phase, ac);
622 return;
623 }
624
625 if (ShenandoahCloneRuntime || !ac->is_clone_array() || !is_oop_array) {
626 // Still looks like an instance? Likely a large instance or reflective
627 // clone with unknown length. Go to runtime and handle it there.
628 clone_in_runtime(phase, ac, ShenandoahRuntime::clone_addr(), "ShenandoahRuntime::clone");
629 return;
630 }
631
632 // We are cloning the oop array. Call into normal oop array copy stubs.
633 // Those stubs would call BarrierSetAssembler to handle GC barriers.
634
635 // This is the full clone, so offsets should equal each other and be at array base.
636 assert(src_offset == dest_offset, "should be equal");
637 const jlong offset = src_offset->get_long();
638 const TypeAryPtr* const ary_ptr = src->get_ptr_type()->isa_aryptr();
639 BasicType bt = ary_ptr->elem()->array_element_basic_type();
640 assert(offset == arrayOopDesc::base_offset_in_bytes(bt), "should match");
641
642 const char* copyfunc_name = "arraycopy";
643 const address copyfunc_addr = phase->basictype2arraycopy(T_OBJECT, nullptr, nullptr, true, copyfunc_name, true);
644
645 Node* const call = phase->make_leaf_call(ctrl, mem,
646 OptoRuntime::fast_arraycopy_Type(),
647 copyfunc_addr, copyfunc_name,
648 TypeRawPtr::BOTTOM,
649 phase->basic_plus_adr(src, src_offset),
650 phase->basic_plus_adr(dest, dest_offset),
651 length,
652 phase->top()
653 );
654 phase->transform_later(call);
655
656 phase->igvn().replace_node(ac, call);
657 }
658
659 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
660 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
661 }
662
663 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
664 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
665 }
666
667 void ShenandoahBarrierSetC2::print_barrier_data(outputStream* os, uint8_t data) {
668 os->print(" Node barriers: ");
669 if ((data & ShenandoahBitStrong) != 0) {
670 data &= ~ShenandoahBitStrong;
671 os->print("strong ");
672 }
673
674 if ((data & ShenandoahBitWeak) != 0) {
675 data &= ~ShenandoahBitWeak;
676 os->print("weak ");
677 }
678
679 if ((data & ShenandoahBitPhantom) != 0) {
680 data &= ~ShenandoahBitPhantom;
681 os->print("phantom ");
682 }
683
684 if ((data & ShenandoahBitKeepAlive) != 0) {
685 data &= ~ShenandoahBitKeepAlive;
686 os->print("keepalive ");
687 }
688
689 if ((data & ShenandoahBitCardMark) != 0) {
690 data &= ~ShenandoahBitCardMark;
691 os->print("cardmark ");
692 }
693
694 if ((data & ShenandoahBitNative) != 0) {
695 data &= ~ShenandoahBitNative;
696 os->print("native ");
697 }
698
699 if ((data & ShenandoahBitNotNull) != 0) {
700 data &= ~ShenandoahBitNotNull;
701 os->print("not-null ");
702 }
703
704 if ((data & ShenandoahBitElided) != 0) {
705 data &= ~ShenandoahBitElided;
706 os->print("elided ");
707 }
708
709 os->cr();
710
711 if (data > 0) {
712 fatal("Unknown bit!");
713 }
714
715 os->print_cr(" GC configuration: %sLRB %sSATB %sCAS %sClone %sCard",
716 (ShenandoahLoadRefBarrier ? "+" : "-"),
717 (ShenandoahSATBBarrier ? "+" : "-"),
718 (ShenandoahCASBarrier ? "+" : "-"),
719 (ShenandoahCloneBarrier ? "+" : "-"),
720 (ShenandoahCardBarrier ? "+" : "-")
721 );
722 }
723
724
725 #ifdef ASSERT
726 void ShenandoahBarrierSetC2::verify_gc_barrier_assert(bool cond, const char* msg, uint8_t bd, Node* n) {
727 if (!cond) {
728 stringStream ss;
729 ss.print_cr("%s", msg);
730 ss.print_cr("-----------------");
731 print_barrier_data(&ss, bd);
732 ss.print_cr("-----------------");
733 n->dump_bfs(1, nullptr, "", &ss);
734 report_vm_error(__FILE__, __LINE__, ss.as_string());
735 }
736 }
737
738 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
739 if (!ShenandoahVerifyOptoBarriers) {
740 return;
741 }
742
743 // Verify depending on the barriers actually enabled, allowing verification in passive mode.
744 // Normally, we have _some_ bits set on all accesses. Optimizations may drop some bits,
745 // but only the last optimization step eliminates all remaining metadata flags. Only then
746 // the access data can be completely blank.
747 bool final_phase = (phase == BeforeCodeGen);
748 bool expect_load_barriers = !final_phase && ShenandoahLoadRefBarrier;
749 bool expect_store_barriers = !final_phase && (ShenandoahSATBBarrier || ShenandoahCardBarrier);
750 bool expect_load_store_barriers = expect_load_barriers || expect_store_barriers;
751 bool expect_some_real = final_phase;
752
753 Unique_Node_List wq;
754
755 RootNode* root = compile->root();
756 wq.push(root);
757
758 // Also seed the outs to capture nodes are not reachable from in()-s, e.g. endless loops.
759 for (DUIterator_Fast imax, i = root->fast_outs(imax); i < imax; i++) {
760 Node* m = root->fast_out(i);
761 wq.push(m);
762 }
763
764 for (uint next = 0; next < wq.size(); next++) {
765 Node *n = wq.at(next);
766 assert(!n->is_Mach(), "No Mach nodes here yet");
767
768 int opc = n->Opcode();
769
770 uint8_t bd = 0;
771 const TypePtr* adr_type = nullptr;
772 if (is_Load(opc)) {
773 bd = n->as_Load()->barrier_data();
774 adr_type = n->as_Load()->adr_type();
775 } else if (is_Store(opc)) {
776 bd = n->as_Store()->barrier_data();
777 adr_type = n->as_Store()->adr_type();
778 } else if (is_LoadStore(opc)) {
779 bd = n->as_LoadStore()->barrier_data();
780 adr_type = n->as_LoadStore()->adr_type();
781 } else if (n->is_Mem()) {
782 bd = MemNode::barrier_data(n);
783 verify_gc_barrier_assert(bd == 0, "Other mem nodes should have no barrier data", bd, n);
784 }
785
786 bool is_weak = (bd & (ShenandoahBitWeak | ShenandoahBitPhantom)) != 0;
787 bool is_native = (bd & ShenandoahBitNative) != 0;
788
789 bool is_referent = adr_type != nullptr &&
790 adr_type->isa_instptr() &&
791 adr_type->is_instptr()->instance_klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
792 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset();
793
794 bool is_oop_addr = (adr_type != nullptr) && (adr_type->isa_oopptr() || adr_type->isa_narrowoop());
795 bool is_raw_addr = (adr_type != nullptr) && (adr_type->isa_rawptr() || adr_type->isa_klassptr());
796
797 verify_gc_barrier_assert(!expect_some_real || (bd == 0) || (bd & ShenandoahBitsReal) != 0, "Without real barriers, metadata should be stripped at this point", bd, n);
798
799 if (is_oop_addr) {
800 if (is_Load(opc)) {
801 verify_gc_barrier_assert(!expect_load_barriers || (bd != 0), "Oop load should have barrier data", bd, n);
802 verify_gc_barrier_assert(!is_weak || is_referent, "Weak load only for Reference.referent", bd, n);
803 } else if (is_Store(opc)) {
804 // Reference.referent stores can be without barriers.
805 verify_gc_barrier_assert(!expect_store_barriers || is_referent || (bd != 0), "Oop store should have barrier data", bd, n);
806 } else if (is_LoadStore(opc)) {
807 verify_gc_barrier_assert(!expect_load_store_barriers || (bd != 0), "Oop load-store should have barrier data", bd, n);
808 }
809 } else if (is_raw_addr) {
810 if (is_native) {
811 if (is_Load(opc)) {
812 verify_gc_barrier_assert(!expect_load_barriers || (bd != 0), "Native oop load should have barrier data", bd, n);
813 }
814 if (is_Store(opc)) {
815 verify_gc_barrier_assert(!expect_store_barriers || (bd != 0), "Native oop store should have barrier data", bd, n);
816 }
817 if (is_LoadStore(opc)) {
818 verify_gc_barrier_assert(!expect_load_store_barriers || (bd != 0), "Native oop load-store should have barrier data", bd, n);
819 }
820 } else {
821 // Some Load/Stores are used for T_ADDRESS and/or raw stores, which are supposed not to have barriers.
822 // Some other Load/Stores are emitted for real oops, but on raw addresses via Unsafe.
823 // The distinction on this level is lost, so we cannot really verify this.
824 }
825 } else {
826 if (is_Load(opc) || is_Store(opc) || is_LoadStore(opc)) {
827 verify_gc_barrier_assert(false, "Unclassified access type", bd, n);
828 }
829 }
830
831 for (uint j = 0; j < n->req(); j++) {
832 Node* in = n->in(j);
833 if (in != nullptr) {
834 wq.push(in);
835 }
836 }
837 }
838 }
839 #endif
840
841 static ShenandoahBarrierSetC2State* barrier_set_state() {
842 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
843 }
844
845 int ShenandoahBarrierSetC2::estimate_stub_size() const {
846 GrowableArray<ShenandoahBarrierStubC2*>* const stubs = barrier_set_state()->stubs();
847 assert(stubs->is_empty(), "Lifecycle: no stubs were yet created");
848 return 0;
849 }
850
851 void ShenandoahBarrierSetC2::emit_stubs(CodeBuffer& cb) const {
852 MacroAssembler masm(&cb);
853
854 PhaseOutput* const output = Compile::current()->output();
855 assert(masm.offset() <= output->buffer_sizing_data()->_code,
856 "Stubs are assumed to be emitted directly after code and code_size is a hard limit on where it can start");
857 barrier_set_state()->set_stubs_start_offset(masm.offset());
858
859 // Stub generation counts all stubs as skipped for the sake of inlining policy.
860 // This is critical for performance, check it.
861 #ifdef ASSERT
862 int offset_before = masm.offset();
863 int skipped_before = cb.total_skipped_instructions_size();
864 #endif
865
866 GrowableArray<ShenandoahBarrierStubC2*>* const stubs = barrier_set_state()->stubs();
867 for (int i = 0; i < stubs->length(); i++) {
868 // Make sure there is enough space in the code buffer
869 if (cb.insts()->maybe_expand_to_ensure_remaining(PhaseOutput::MAX_inst_size) && cb.blob() == nullptr) {
870 ciEnv::current()->record_failure("CodeCache is full");
871 return;
872 }
873 stubs->at(i)->emit_code(masm);
874 }
875
876 #ifdef ASSERT
877 int offset_after = masm.offset();
878 int skipped_after = cb.total_skipped_instructions_size();
879 assert(offset_after - offset_before == skipped_after - skipped_before,
880 "All stubs are counted as skipped. masm: %d - %d = %d, cb: %d - %d = %d",
881 offset_after, offset_before, offset_after - offset_before,
882 skipped_after, skipped_before, skipped_after - skipped_before);
883 #endif
884
885 masm.flush();
886 }
887
888 void ShenandoahBarrierStubC2::register_stub(ShenandoahBarrierStubC2* stub) {
889 if (!Compile::current()->output()->in_scratch_emit_size()) {
890 barrier_set_state()->stubs()->append(stub);
891 }
892 }
893
894 ShenandoahBarrierStubC2* ShenandoahBarrierStubC2::create(const MachNode* node, Register obj, Address addr, Register tmp1, Register tmp2, bool narrow, bool do_load) {
895 auto* stub = new (Compile::current()->comp_arena()) ShenandoahBarrierStubC2(node, obj, addr, tmp1, tmp2, narrow, do_load);
896 register_stub(stub);
897 return stub;
898 }
899
900 void ShenandoahBarrierStubC2::load_post(MacroAssembler* masm, const MachNode* node, Register obj, Address addr, Register tmp1, Register tmp2, bool narrow) {
901 // Load post-barrier:
902 // a. Satisfies the need for LRB for normal loads
903 // b. Passes a weak load through LRB-weak
904 // c. Keep-alives a weak load
905 if (needs_slow_barrier(node)) {
906 ShenandoahBarrierStubC2* const stub = create(node, obj, addr, tmp1, tmp2, narrow, /* do_load = */ false);
907 char check = 0;
908 check |= needs_keep_alive_barrier(node) ? ShenandoahHeap::MARKING : 0;
909 check |= needs_load_ref_barrier(node) ? ShenandoahHeap::HAS_FORWARDED : 0;
910 check |= needs_load_ref_barrier_weak(node) ? ShenandoahHeap::WEAK_ROOTS : 0;
911 stub->enter_if_gc_state(*masm, check, tmp1);
912 }
913 }
914
915 void ShenandoahBarrierStubC2::store_pre(MacroAssembler* masm, const MachNode* node, Register obj, Address addr, Register tmp1, Register tmp2, bool narrow) {
916 // Store pre-barrier: SATB, keep-alive the current memory value.
917 if (needs_slow_barrier(node)) {
918 assert(!needs_load_ref_barrier(node), "Should not be required for stores");
919 ShenandoahBarrierStubC2* const stub = create(node, obj, addr, tmp1, tmp2, narrow, /* do_load = */ true);
920 stub->enter_if_gc_state(*masm, ShenandoahHeap::MARKING, tmp1);
921 }
922 }
923
924 void ShenandoahBarrierStubC2::load_store_pre(MacroAssembler* masm, const MachNode* node, Register obj, Address addr, Register tmp1, Register tmp2, bool narrow) {
925 // Load/Store pre-barrier:
926 // a. Avoids false positives from CAS encountering to-space memory values.
927 // b. Satisfies the need for LRB for the CAE result.
928 // c. Records old value for the sake of SATB.
929 //
930 // (a) and (b) are covered because load barrier does memory location fixup.
931 // (c) is covered by KA on the current memory value.
932 if (needs_slow_barrier(node)) {
933 ShenandoahBarrierStubC2* const stub = create(node, obj, addr, tmp1, tmp2, narrow, /* do_load = */ true);
934 char check = 0;
935 check |= needs_keep_alive_barrier(node) ? ShenandoahHeap::MARKING : 0;
936 check |= needs_load_ref_barrier(node) ? ShenandoahHeap::HAS_FORWARDED : 0;
937 assert(!needs_load_ref_barrier_weak(node), "Not supported for Load/Stores");
938 stub->enter_if_gc_state(*masm, check, tmp1);
939 }
940 }
941
942 void ShenandoahBarrierStubC2::store_post(MacroAssembler* masm, const MachNode* node, Address addr, Register tmp1, Register tmp2) {
943 if (needs_card_barrier(node)) {
944 cardtable(*masm, addr, tmp1, tmp2);
945 }
946 }
947
948 void ShenandoahBarrierStubC2::load_store_post(MacroAssembler* masm, const MachNode* node, Address addr, Register tmp1, Register tmp2) {
949 store_post(masm, node, addr, tmp1, tmp2);
950 }
951
952 bool ShenandoahBarrierStubC2::is_live_register(Register reg) {
953 return preserve_set().member(OptoReg::as_OptoReg(reg->as_VMReg()));
954 }
955
956 Register ShenandoahBarrierStubC2::select_temp_register(bool& selected_live, Register skip_reg1, Register skip_reg2) {
957 Register tmp = noreg;
958 Register fallback_live = noreg;
959
960 // Try to select non-live first:
961 for (int i = 0; i < available_gp_registers(); i++) {
962 Register r = as_Register(i);
963 if (r != _obj && r != _addr.base() && r != _addr.index() &&
964 r != skip_reg1 && r != skip_reg2 && !is_special_register(r)) {
965 if (!is_live_register(r)) {
966 tmp = r;
967 break;
968 } else if (fallback_live == noreg) {
969 fallback_live = r;
970 }
971 }
972 }
973
974 // If we could not find a non-live register, select the live fallback:
975 if (tmp == noreg) {
976 tmp = fallback_live;
977 selected_live = true;
978 } else {
979 selected_live = false;
980 }
981
982 assert(tmp != noreg, "successfully selected");
983 assert_different_registers(tmp, skip_reg1);
984 assert_different_registers(tmp, skip_reg2);
985 assert_different_registers(tmp, _obj);
986 assert_different_registers(tmp, _addr.base());
987 assert_different_registers(tmp, _addr.index());
988 return tmp;
989 }
990
991 address ShenandoahBarrierStubC2::keepalive_runtime_entry_addr() {
992 if (_narrow) {
993 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre_narrow);
994 } else {
995 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre);
996 }
997 }
998
999 address ShenandoahBarrierStubC2::lrb_runtime_entry_addr() {
1000 bool is_strong = (_node->barrier_data() & ShenandoahBitStrong) != 0;
1001 bool is_weak = (_node->barrier_data() & ShenandoahBitWeak) != 0;
1002 bool is_phantom = (_node->barrier_data() & ShenandoahBitPhantom) != 0;
1003
1004 if (_narrow) {
1005 if (is_strong) {
1006 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow_narrow);
1007 } else if (is_weak) {
1008 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow_narrow);
1009 } else if (is_phantom) {
1010 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow_narrow);
1011 }
1012 } else {
1013 if (is_strong) {
1014 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
1015 } else if (is_weak) {
1016 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
1017 } else if (is_phantom) {
1018 return CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
1019 }
1020 }
1021
1022 ShouldNotReachHere();
1023 return nullptr;
1024 }
1025
1026 bool ShenandoahBarrierSetC2State::needs_liveness_data(const MachNode* mach) const {
1027 // Nodes that require slow-path stubs need liveness data.
1028 return ShenandoahBarrierStubC2::needs_slow_barrier(mach);
1029 }
1030
1031 bool ShenandoahBarrierSetC2State::needs_livein_data() const {
1032 return true;
1033 }
|