1 /* 2 * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved. 3 * 4 * This code is free software; you can redistribute it and/or modify it 5 * under the terms of the GNU General Public License version 2 only, as 6 * published by the Free Software Foundation. 7 * 8 * This code is distributed in the hope that it will be useful, but WITHOUT 9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 11 * version 2 for more details (a copy is included in the LICENSE file that 12 * accompanied this code). 13 * 14 * You should have received a copy of the GNU General Public License version 15 * 2 along with this work; if not, write to the Free Software Foundation, 16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 17 * 18 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 19 * or visit www.oracle.com if you need additional information or have any 20 * questions. 21 * 22 */ 23 24 #include "precompiled.hpp" 25 #include "gc/shared/barrierSet.hpp" 26 #include "gc/shenandoah/shenandoahForwarding.hpp" 27 #include "gc/shenandoah/shenandoahHeap.hpp" 28 #include "gc/shenandoah/shenandoahHeuristics.hpp" 29 #include "gc/shenandoah/shenandoahRuntime.hpp" 30 #include "gc/shenandoah/shenandoahThreadLocalData.hpp" 31 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp" 32 #include "gc/shenandoah/c2/shenandoahSupport.hpp" 33 #include "opto/arraycopynode.hpp" 34 #include "opto/escape.hpp" 35 #include "opto/graphKit.hpp" 36 #include "opto/idealKit.hpp" 37 #include "opto/macro.hpp" 38 #include "opto/movenode.hpp" 39 #include "opto/narrowptrnode.hpp" 40 #include "opto/rootnode.hpp" 41 42 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() { 43 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2()); 44 } 45 46 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena) 47 : _enqueue_barriers(new (comp_arena) GrowableArray<ShenandoahEnqueueBarrierNode*>(comp_arena, 8, 0, NULL)), 48 _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8, 0, NULL)) { 49 } 50 51 int ShenandoahBarrierSetC2State::enqueue_barriers_count() const { 52 return _enqueue_barriers->length(); 53 } 54 55 ShenandoahEnqueueBarrierNode* ShenandoahBarrierSetC2State::enqueue_barrier(int idx) const { 56 return _enqueue_barriers->at(idx); 57 } 58 59 void ShenandoahBarrierSetC2State::add_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) { 60 assert(!_enqueue_barriers->contains(n), "duplicate entry in barrier list"); 61 _enqueue_barriers->append(n); 62 } 63 64 void ShenandoahBarrierSetC2State::remove_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) { 65 if (_enqueue_barriers->contains(n)) { 66 _enqueue_barriers->remove(n); 67 } 68 } 69 70 int ShenandoahBarrierSetC2State::load_reference_barriers_count() const { 71 return _load_reference_barriers->length(); 72 } 73 74 ShenandoahLoadReferenceBarrierNode* ShenandoahBarrierSetC2State::load_reference_barrier(int idx) const { 75 return _load_reference_barriers->at(idx); 76 } 77 78 void ShenandoahBarrierSetC2State::add_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) { 79 assert(!_load_reference_barriers->contains(n), "duplicate entry in barrier list"); 80 _load_reference_barriers->append(n); 81 } 82 83 void ShenandoahBarrierSetC2State::remove_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) { 84 if (_load_reference_barriers->contains(n)) { 85 _load_reference_barriers->remove(n); 86 } 87 } 88 89 Node* ShenandoahBarrierSetC2::shenandoah_storeval_barrier(GraphKit* kit, Node* obj) const { 90 if (ShenandoahStoreValEnqueueBarrier) { 91 obj = shenandoah_enqueue_barrier(kit, obj); 92 } 93 return obj; 94 } 95 96 #define __ kit-> 97 98 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseTransform* phase, Node* adr, 99 BasicType bt, uint adr_idx) const { 100 intptr_t offset = 0; 101 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 102 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 103 104 if (offset == Type::OffsetBot) { 105 return false; // cannot unalias unless there are precise offsets 106 } 107 108 if (alloc == NULL) { 109 return false; // No allocation found 110 } 111 112 intptr_t size_in_bytes = type2aelembytes(bt); 113 114 Node* mem = __ memory(adr_idx); // start searching here... 115 116 for (int cnt = 0; cnt < 50; cnt++) { 117 118 if (mem->is_Store()) { 119 120 Node* st_adr = mem->in(MemNode::Address); 121 intptr_t st_offset = 0; 122 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 123 124 if (st_base == NULL) { 125 break; // inscrutable pointer 126 } 127 128 // Break we have found a store with same base and offset as ours so break 129 if (st_base == base && st_offset == offset) { 130 break; 131 } 132 133 if (st_offset != offset && st_offset != Type::OffsetBot) { 134 const int MAX_STORE = BytesPerLong; 135 if (st_offset >= offset + size_in_bytes || 136 st_offset <= offset - MAX_STORE || 137 st_offset <= offset - mem->as_Store()->memory_size()) { 138 // Success: The offsets are provably independent. 139 // (You may ask, why not just test st_offset != offset and be done? 140 // The answer is that stores of different sizes can co-exist 141 // in the same sequence of RawMem effects. We sometimes initialize 142 // a whole 'tile' of array elements with a single jint or jlong.) 143 mem = mem->in(MemNode::Memory); 144 continue; // advance through independent store memory 145 } 146 } 147 148 if (st_base != base 149 && MemNode::detect_ptr_independence(base, alloc, st_base, 150 AllocateNode::Ideal_allocation(st_base, phase), 151 phase)) { 152 // Success: The bases are provably independent. 153 mem = mem->in(MemNode::Memory); 154 continue; // advance through independent store memory 155 } 156 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 157 158 InitializeNode* st_init = mem->in(0)->as_Initialize(); 159 AllocateNode* st_alloc = st_init->allocation(); 160 161 // Make sure that we are looking at the same allocation site. 162 // The alloc variable is guaranteed to not be null here from earlier check. 163 if (alloc == st_alloc) { 164 // Check that the initialization is storing NULL so that no previous store 165 // has been moved up and directly write a reference 166 Node* captured_store = st_init->find_captured_store(offset, 167 type2aelembytes(T_OBJECT), 168 phase); 169 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 170 return true; 171 } 172 } 173 } 174 175 // Unless there is an explicit 'continue', we must bail out here, 176 // because 'mem' is an inscrutable memory state (e.g., a call). 177 break; 178 } 179 180 return false; 181 } 182 183 #undef __ 184 #define __ ideal. 185 186 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit, 187 bool do_load, 188 Node* obj, 189 Node* adr, 190 uint alias_idx, 191 Node* val, 192 const TypeOopPtr* val_type, 193 Node* pre_val, 194 BasicType bt) const { 195 // Some sanity checks 196 // Note: val is unused in this routine. 197 198 if (do_load) { 199 // We need to generate the load of the previous value 200 assert(obj != NULL, "must have a base"); 201 assert(adr != NULL, "where are loading from?"); 202 assert(pre_val == NULL, "loaded already?"); 203 assert(val_type != NULL, "need a type"); 204 205 if (ReduceInitialCardMarks 206 && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) { 207 return; 208 } 209 210 } else { 211 // In this case both val_type and alias_idx are unused. 212 assert(pre_val != NULL, "must be loaded already"); 213 // Nothing to be done if pre_val is null. 214 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 215 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 216 } 217 assert(bt == T_OBJECT, "or we shouldn't be here"); 218 219 IdealKit ideal(kit, true); 220 221 Node* tls = __ thread(); // ThreadLocalStorage 222 223 Node* no_base = __ top(); 224 Node* zero = __ ConI(0); 225 Node* zeroX = __ ConX(0); 226 227 float likely = PROB_LIKELY(0.999); 228 float unlikely = PROB_UNLIKELY(0.999); 229 230 // Offsets into the thread 231 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()); 232 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()); 233 234 // Now the actual pointers into the thread 235 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 236 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 237 238 // Now some of the values 239 Node* marking; 240 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()))); 241 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw); 242 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING)); 243 assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape"); 244 245 // if (!marking) 246 __ if_then(marking, BoolTest::ne, zero, unlikely); { 247 BasicType index_bt = TypeX_X->basic_type(); 248 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 249 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 250 251 if (do_load) { 252 // load original value 253 // alias_idx correct?? 254 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 255 } 256 257 // if (pre_val != NULL) 258 __ if_then(pre_val, BoolTest::ne, kit->null()); { 259 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 260 261 // is the queue for this thread full? 262 __ if_then(index, BoolTest::ne, zeroX, likely); { 263 264 // decrement the index 265 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 266 267 // Now get the buffer location we will log the previous value into and store it 268 Node *log_addr = __ AddP(no_base, buffer, next_index); 269 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 270 // update the index 271 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 272 273 } __ else_(); { 274 275 // logging buffer is full, call the runtime 276 const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type(); 277 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls); 278 } __ end_if(); // (!index) 279 } __ end_if(); // (pre_val != NULL) 280 } __ end_if(); // (!marking) 281 282 // Final sync IdealKit and GraphKit. 283 kit->final_sync(ideal); 284 285 if (ShenandoahSATBBarrier && adr != NULL) { 286 Node* c = kit->control(); 287 Node* call = c->in(1)->in(1)->in(1)->in(0); 288 assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected"); 289 call->add_req(adr); 290 } 291 } 292 293 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) { 294 return call->is_CallLeaf() && 295 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry); 296 } 297 298 bool ShenandoahBarrierSetC2::is_shenandoah_lrb_call(Node* call) { 299 return call->is_CallLeaf() && 300 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier); 301 } 302 303 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseTransform *phase, Node* n) { 304 if (n->Opcode() != Op_If) { 305 return false; 306 } 307 308 Node* bol = n->in(1); 309 assert(bol->is_Bool(), ""); 310 Node* cmpx = bol->in(1); 311 if (bol->as_Bool()->_test._test == BoolTest::ne && 312 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) && 313 is_shenandoah_state_load(cmpx->in(1)->in(1)) && 314 cmpx->in(1)->in(2)->is_Con() && 315 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) { 316 return true; 317 } 318 319 return false; 320 } 321 322 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) { 323 if (!n->is_Load()) return false; 324 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset()); 325 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal 326 && n->in(2)->in(3)->is_Con() 327 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset; 328 } 329 330 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit, 331 bool do_load, 332 Node* obj, 333 Node* adr, 334 uint alias_idx, 335 Node* val, 336 const TypeOopPtr* val_type, 337 Node* pre_val, 338 BasicType bt) const { 339 if (ShenandoahSATBBarrier) { 340 IdealKit ideal(kit); 341 kit->sync_kit(ideal); 342 343 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt); 344 345 ideal.sync_kit(kit); 346 kit->final_sync(ideal); 347 } 348 } 349 350 Node* ShenandoahBarrierSetC2::shenandoah_enqueue_barrier(GraphKit* kit, Node* pre_val) const { 351 return kit->gvn().transform(new ShenandoahEnqueueBarrierNode(pre_val)); 352 } 353 354 // Helper that guards and inserts a pre-barrier. 355 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset, 356 Node* pre_val, bool need_mem_bar) const { 357 // We could be accessing the referent field of a reference object. If so, when G1 358 // is enabled, we need to log the value in the referent field in an SATB buffer. 359 // This routine performs some compile time filters and generates suitable 360 // runtime filters that guard the pre-barrier code. 361 // Also add memory barrier for non volatile load from the referent field 362 // to prevent commoning of loads across safepoint. 363 364 // Some compile time checks. 365 366 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 367 const TypeX* otype = offset->find_intptr_t_type(); 368 if (otype != NULL && otype->is_con() && 369 otype->get_con() != java_lang_ref_Reference::referent_offset) { 370 // Constant offset but not the reference_offset so just return 371 return; 372 } 373 374 // We only need to generate the runtime guards for instances. 375 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 376 if (btype != NULL) { 377 if (btype->isa_aryptr()) { 378 // Array type so nothing to do 379 return; 380 } 381 382 const TypeInstPtr* itype = btype->isa_instptr(); 383 if (itype != NULL) { 384 // Can the klass of base_oop be statically determined to be 385 // _not_ a sub-class of Reference and _not_ Object? 386 ciKlass* klass = itype->klass(); 387 if ( klass->is_loaded() && 388 !klass->is_subtype_of(kit->env()->Reference_klass()) && 389 !kit->env()->Object_klass()->is_subtype_of(klass)) { 390 return; 391 } 392 } 393 } 394 395 // The compile time filters did not reject base_oop/offset so 396 // we need to generate the following runtime filters 397 // 398 // if (offset == java_lang_ref_Reference::_reference_offset) { 399 // if (instance_of(base, java.lang.ref.Reference)) { 400 // pre_barrier(_, pre_val, ...); 401 // } 402 // } 403 404 float likely = PROB_LIKELY( 0.999); 405 float unlikely = PROB_UNLIKELY(0.999); 406 407 IdealKit ideal(kit); 408 409 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 410 411 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 412 // Update graphKit memory and control from IdealKit. 413 kit->sync_kit(ideal); 414 415 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass())); 416 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con); 417 418 // Update IdealKit memory and control from graphKit. 419 __ sync_kit(kit); 420 421 Node* one = __ ConI(1); 422 // is_instof == 0 if base_oop == NULL 423 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 424 425 // Update graphKit from IdeakKit. 426 kit->sync_kit(ideal); 427 428 // Use the pre-barrier to record the value in the referent field 429 satb_write_barrier_pre(kit, false /* do_load */, 430 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 431 pre_val /* pre_val */, 432 T_OBJECT); 433 if (need_mem_bar) { 434 // Add memory barrier to prevent commoning reads from this field 435 // across safepoint since GC can change its value. 436 kit->insert_mem_bar(Op_MemBarCPUOrder); 437 } 438 // Update IdealKit from graphKit. 439 __ sync_kit(kit); 440 441 } __ end_if(); // _ref_type != ref_none 442 } __ end_if(); // offset == referent_offset 443 444 // Final sync IdealKit and GraphKit. 445 kit->final_sync(ideal); 446 } 447 448 #undef __ 449 450 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() { 451 const Type **fields = TypeTuple::fields(2); 452 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value 453 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread 454 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields); 455 456 // create result type (range) 457 fields = TypeTuple::fields(0); 458 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields); 459 460 return TypeFunc::make(domain, range); 461 } 462 463 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() { 464 const Type **fields = TypeTuple::fields(1); 465 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value 466 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields); 467 468 // create result type (range) 469 fields = TypeTuple::fields(0); 470 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields); 471 472 return TypeFunc::make(domain, range); 473 } 474 475 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type() { 476 const Type **fields = TypeTuple::fields(1); 477 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value 478 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields); 479 480 // create result type (range) 481 fields = TypeTuple::fields(1); 482 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; 483 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields); 484 485 return TypeFunc::make(domain, range); 486 } 487 488 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const { 489 DecoratorSet decorators = access.decorators(); 490 491 const TypePtr* adr_type = access.addr().type(); 492 Node* adr = access.addr().node(); 493 494 bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0; 495 bool on_heap = (decorators & IN_HEAP) != 0; 496 497 if (!access.is_oop() || (!on_heap && !anonymous)) { 498 return BarrierSetC2::store_at_resolved(access, val); 499 } 500 501 GraphKit* kit = access.kit(); 502 503 uint adr_idx = kit->C->get_alias_index(adr_type); 504 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 505 Node* value = val.node(); 506 value = shenandoah_storeval_barrier(kit, value); 507 val.set_node(value); 508 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(), 509 static_cast<const TypeOopPtr*>(val.type()), NULL /* pre_val */, access.type()); 510 return BarrierSetC2::store_at_resolved(access, val); 511 } 512 513 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const { 514 DecoratorSet decorators = access.decorators(); 515 516 Node* adr = access.addr().node(); 517 Node* obj = access.base(); 518 519 bool mismatched = (decorators & C2_MISMATCHED) != 0; 520 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0; 521 bool on_heap = (decorators & IN_HEAP) != 0; 522 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0; 523 bool is_unordered = (decorators & MO_UNORDERED) != 0; 524 bool need_cpu_mem_bar = !is_unordered || mismatched || !on_heap; 525 526 Node* top = Compile::current()->top(); 527 528 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top; 529 Node* load = BarrierSetC2::load_at_resolved(access, val_type); 530 531 if (access.is_oop()) { 532 if (ShenandoahLoadRefBarrier) { 533 load = new ShenandoahLoadReferenceBarrierNode(NULL, load); 534 load = access.kit()->gvn().transform(load); 535 } 536 } 537 538 // If we are reading the value of the referent field of a Reference 539 // object (either by using Unsafe directly or through reflection) 540 // then, if SATB is enabled, we need to record the referent in an 541 // SATB log buffer using the pre-barrier mechanism. 542 // Also we need to add memory barrier to prevent commoning reads 543 // from this field across safepoint since GC can change its value. 544 bool need_read_barrier = ShenandoahKeepAliveBarrier && 545 (on_heap && (on_weak || (unknown && offset != top && obj != top))); 546 547 if (!access.is_oop() || !need_read_barrier) { 548 return load; 549 } 550 551 GraphKit* kit = access.kit(); 552 553 if (on_weak) { 554 // Use the pre-barrier to record the value in the referent field 555 satb_write_barrier_pre(kit, false /* do_load */, 556 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 557 load /* pre_val */, T_OBJECT); 558 // Add memory barrier to prevent commoning reads from this field 559 // across safepoint since GC can change its value. 560 kit->insert_mem_bar(Op_MemBarCPUOrder); 561 } else if (unknown) { 562 // We do not require a mem bar inside pre_barrier if need_mem_bar 563 // is set: the barriers would be emitted by us. 564 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar); 565 } 566 567 return load; 568 } 569 570 static void pin_atomic_op(C2AtomicAccess& access) { 571 if (!access.needs_pinning()) { 572 return; 573 } 574 // SCMemProjNodes represent the memory state of a LoadStore. Their 575 // main role is to prevent LoadStore nodes from being optimized away 576 // when their results aren't used. 577 GraphKit* kit = access.kit(); 578 Node* load_store = access.raw_access(); 579 assert(load_store != NULL, "must pin atomic op"); 580 Node* proj = kit->gvn().transform(new SCMemProjNode(load_store)); 581 kit->set_memory(proj, access.alias_idx()); 582 } 583 584 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicAccess& access, Node* expected_val, 585 Node* new_val, const Type* value_type) const { 586 GraphKit* kit = access.kit(); 587 if (access.is_oop()) { 588 new_val = shenandoah_storeval_barrier(kit, new_val); 589 shenandoah_write_barrier_pre(kit, false /* do_load */, 590 NULL, NULL, max_juint, NULL, NULL, 591 expected_val /* pre_val */, T_OBJECT); 592 593 MemNode::MemOrd mo = access.mem_node_mo(); 594 Node* mem = access.memory(); 595 Node* adr = access.addr().node(); 596 const TypePtr* adr_type = access.addr().type(); 597 Node* load_store = NULL; 598 599 #ifdef _LP64 600 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 601 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop())); 602 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop())); 603 if (ShenandoahCASBarrier) { 604 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo)); 605 } else { 606 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo)); 607 } 608 } else 609 #endif 610 { 611 if (ShenandoahCASBarrier) { 612 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo)); 613 } else { 614 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo)); 615 } 616 } 617 618 access.set_raw_access(load_store); 619 pin_atomic_op(access); 620 621 #ifdef _LP64 622 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 623 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 624 } 625 #endif 626 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, load_store)); 627 return load_store; 628 } 629 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type); 630 } 631 632 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicAccess& access, Node* expected_val, 633 Node* new_val, const Type* value_type) const { 634 GraphKit* kit = access.kit(); 635 if (access.is_oop()) { 636 new_val = shenandoah_storeval_barrier(kit, new_val); 637 shenandoah_write_barrier_pre(kit, false /* do_load */, 638 NULL, NULL, max_juint, NULL, NULL, 639 expected_val /* pre_val */, T_OBJECT); 640 DecoratorSet decorators = access.decorators(); 641 MemNode::MemOrd mo = access.mem_node_mo(); 642 Node* mem = access.memory(); 643 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0; 644 Node* load_store = NULL; 645 Node* adr = access.addr().node(); 646 #ifdef _LP64 647 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 648 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop())); 649 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop())); 650 if (ShenandoahCASBarrier) { 651 if (is_weak_cas) { 652 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 653 } else { 654 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 655 } 656 } else { 657 if (is_weak_cas) { 658 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 659 } else { 660 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 661 } 662 } 663 } else 664 #endif 665 { 666 if (ShenandoahCASBarrier) { 667 if (is_weak_cas) { 668 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 669 } else { 670 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 671 } 672 } else { 673 if (is_weak_cas) { 674 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 675 } else { 676 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 677 } 678 } 679 } 680 access.set_raw_access(load_store); 681 pin_atomic_op(access); 682 return load_store; 683 } 684 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type); 685 } 686 687 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicAccess& access, Node* val, const Type* value_type) const { 688 GraphKit* kit = access.kit(); 689 if (access.is_oop()) { 690 val = shenandoah_storeval_barrier(kit, val); 691 } 692 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type); 693 if (access.is_oop()) { 694 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, result)); 695 shenandoah_write_barrier_pre(kit, false /* do_load */, 696 NULL, NULL, max_juint, NULL, NULL, 697 result /* pre_val */, T_OBJECT); 698 } 699 return result; 700 } 701 702 void ShenandoahBarrierSetC2::clone(GraphKit* kit, Node* src, Node* dst, Node* size, bool is_array) const { 703 // TODO: Implement using proper barriers. 704 BarrierSetC2::clone(kit, src, dst, size, is_array); 705 } 706 707 // Support for GC barriers emitted during parsing 708 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const { 709 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) return true; 710 if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) { 711 return false; 712 } 713 CallLeafNode *call = node->as_CallLeaf(); 714 if (call->_name == NULL) { 715 return false; 716 } 717 718 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 || 719 strcmp(call->_name, "shenandoah_cas_obj") == 0 || 720 strcmp(call->_name, "shenandoah_wb_pre") == 0; 721 } 722 723 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const { 724 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) { 725 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn); 726 } 727 if (c->Opcode() == Op_ShenandoahEnqueueBarrier) { 728 c = c->in(1); 729 } 730 return c; 731 } 732 733 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const { 734 return !ShenandoahBarrierC2Support::expand(C, igvn); 735 } 736 737 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const { 738 if (mode == LoopOptsShenandoahExpand) { 739 assert(UseShenandoahGC, "only for shenandoah"); 740 ShenandoahBarrierC2Support::pin_and_expand(phase); 741 return true; 742 } else if (mode == LoopOptsShenandoahPostExpand) { 743 assert(UseShenandoahGC, "only for shenandoah"); 744 visited.Clear(); 745 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase); 746 return true; 747 } 748 return false; 749 } 750 751 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(BasicType type) const { 752 return false; 753 } 754 755 bool ShenandoahBarrierSetC2::clone_needs_postbarrier(ArrayCopyNode *ac, PhaseIterGVN& igvn) { 756 Node* src = ac->in(ArrayCopyNode::Src); 757 const TypeOopPtr* src_type = igvn.type(src)->is_oopptr(); 758 if (src_type->isa_instptr() != NULL) { 759 ciInstanceKlass* ik = src_type->klass()->as_instance_klass(); 760 if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) { 761 if (ik->has_object_fields()) { 762 return true; 763 } else { 764 if (!src_type->klass_is_exact()) { 765 igvn.C->dependencies()->assert_leaf_type(ik); 766 } 767 } 768 } else { 769 return true; 770 } 771 } else if (src_type->isa_aryptr()) { 772 BasicType src_elem = src_type->klass()->as_array_klass()->element_type()->basic_type(); 773 if (src_elem == T_OBJECT || src_elem == T_ARRAY) { 774 return true; 775 } 776 } else { 777 return true; 778 } 779 return false; 780 } 781 782 // Support for macro expanded GC barriers 783 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const { 784 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) { 785 state()->add_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node); 786 } 787 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) { 788 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node); 789 } 790 } 791 792 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const { 793 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) { 794 state()->remove_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node); 795 } 796 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) { 797 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node); 798 } 799 } 800 801 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const { 802 if (is_shenandoah_wb_pre_call(n)) { 803 shenandoah_eliminate_wb_pre(n, ¯o->igvn()); 804 } 805 } 806 807 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const { 808 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), ""); 809 Node* c = call->as_Call()->proj_out(TypeFunc::Control); 810 c = c->unique_ctrl_out(); 811 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 812 c = c->unique_ctrl_out(); 813 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 814 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0); 815 assert(iff->is_If(), "expect test"); 816 if (!is_shenandoah_marking_if(igvn, iff)) { 817 c = c->unique_ctrl_out(); 818 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 819 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0); 820 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test"); 821 } 822 Node* cmpx = iff->in(1)->in(1); 823 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ)); 824 igvn->rehash_node_delayed(call); 825 call->del_req(call->req()-1); 826 } 827 828 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(Unique_Node_List &worklist, Node* node) const { 829 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) { 830 worklist.push(node); 831 } 832 } 833 834 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful) const { 835 for (uint i = 0; i < useful.size(); i++) { 836 Node* n = useful.at(i); 837 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) { 838 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 839 Compile::current()->record_for_igvn(n->fast_out(i)); 840 } 841 } 842 } 843 for (int i = state()->enqueue_barriers_count() - 1; i >= 0; i--) { 844 ShenandoahEnqueueBarrierNode* n = state()->enqueue_barrier(i); 845 if (!useful.member(n)) { 846 state()->remove_enqueue_barrier(n); 847 } 848 } 849 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) { 850 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i); 851 if (!useful.member(n)) { 852 state()->remove_load_reference_barrier(n); 853 } 854 } 855 } 856 857 void ShenandoahBarrierSetC2::add_users_to_worklist(Unique_Node_List* worklist) const {} 858 859 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const { 860 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena); 861 } 862 863 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const { 864 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state()); 865 } 866 867 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be 868 // expanded later, then now is the time to do so. 869 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; } 870 871 #ifdef ASSERT 872 void ShenandoahBarrierSetC2::verify_gc_barriers(bool post_parse) const { 873 if (ShenandoahVerifyOptoBarriers && !post_parse) { 874 ShenandoahBarrierC2Support::verify(Compile::current()->root()); 875 } 876 } 877 #endif 878 879 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const { 880 if (is_shenandoah_wb_pre_call(n)) { 881 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt(); 882 if (n->req() > cnt) { 883 Node* addp = n->in(cnt); 884 if (has_only_shenandoah_wb_pre_uses(addp)) { 885 n->del_req(cnt); 886 if (can_reshape) { 887 phase->is_IterGVN()->_worklist.push(addp); 888 } 889 return n; 890 } 891 } 892 } 893 if (n->Opcode() == Op_CmpP) { 894 Node* in1 = n->in(1); 895 Node* in2 = n->in(2); 896 if (in1->bottom_type() == TypePtr::NULL_PTR) { 897 in2 = step_over_gc_barrier(in2); 898 } 899 if (in2->bottom_type() == TypePtr::NULL_PTR) { 900 in1 = step_over_gc_barrier(in1); 901 } 902 PhaseIterGVN* igvn = phase->is_IterGVN(); 903 if (in1 != n->in(1)) { 904 if (igvn != NULL) { 905 n->set_req_X(1, in1, igvn); 906 } else { 907 n->set_req(1, in1); 908 } 909 assert(in2 == n->in(2), "only one change"); 910 return n; 911 } 912 if (in2 != n->in(2)) { 913 if (igvn != NULL) { 914 n->set_req_X(2, in2, igvn); 915 } else { 916 n->set_req(2, in2); 917 } 918 return n; 919 } 920 } else if (can_reshape && 921 n->Opcode() == Op_If && 922 ShenandoahBarrierC2Support::is_heap_stable_test(n) && 923 n->in(0) != NULL) { 924 Node* dom = n->in(0); 925 Node* prev_dom = n; 926 int op = n->Opcode(); 927 int dist = 16; 928 // Search up the dominator tree for another heap stable test 929 while (dom->Opcode() != op || // Not same opcode? 930 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1? 931 prev_dom->in(0) != dom) { // One path of test does not dominate? 932 if (dist < 0) return NULL; 933 934 dist--; 935 prev_dom = dom; 936 dom = IfNode::up_one_dom(dom); 937 if (!dom) return NULL; 938 } 939 940 // Check that we did not follow a loop back to ourselves 941 if (n == dom) { 942 return NULL; 943 } 944 945 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN()); 946 } 947 return NULL; 948 } 949 950 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) { 951 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 952 Node* u = n->fast_out(i); 953 if (!is_shenandoah_wb_pre_call(u)) { 954 return false; 955 } 956 } 957 return n->outcnt() > 0; 958 }