1 /* 2 * Copyright (c) 2018, 2023, Red Hat, Inc. All rights reserved. 3 * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved. 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 5 * 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 "precompiled.hpp" 27 #include "classfile/javaClasses.hpp" 28 #include "gc/shared/barrierSet.hpp" 29 #include "gc/shenandoah/shenandoahBarrierSet.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 "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp" 35 #include "gc/shenandoah/c2/shenandoahSupport.hpp" 36 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp" 37 #include "gc/shenandoah/mode/shenandoahMode.hpp" 38 #include "opto/arraycopynode.hpp" 39 #include "opto/escape.hpp" 40 #include "opto/graphKit.hpp" 41 #include "opto/idealKit.hpp" 42 #include "opto/macro.hpp" 43 #include "opto/movenode.hpp" 44 #include "opto/narrowptrnode.hpp" 45 #include "opto/rootnode.hpp" 46 #include "opto/runtime.hpp" 47 48 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() { 49 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2()); 50 } 51 52 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena) 53 : _iu_barriers(new (comp_arena) GrowableArray<ShenandoahIUBarrierNode*>(comp_arena, 8, 0, nullptr)), 54 _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8, 0, nullptr)) { 55 } 56 57 int ShenandoahBarrierSetC2State::iu_barriers_count() const { 58 return _iu_barriers->length(); 59 } 60 61 ShenandoahIUBarrierNode* ShenandoahBarrierSetC2State::iu_barrier(int idx) const { 62 return _iu_barriers->at(idx); 63 } 64 65 void ShenandoahBarrierSetC2State::add_iu_barrier(ShenandoahIUBarrierNode* n) { 66 assert(!_iu_barriers->contains(n), "duplicate entry in barrier list"); 67 _iu_barriers->append(n); 68 } 69 70 void ShenandoahBarrierSetC2State::remove_iu_barrier(ShenandoahIUBarrierNode* n) { 71 _iu_barriers->remove_if_existing(n); 72 } 73 74 int ShenandoahBarrierSetC2State::load_reference_barriers_count() const { 75 return _load_reference_barriers->length(); 76 } 77 78 ShenandoahLoadReferenceBarrierNode* ShenandoahBarrierSetC2State::load_reference_barrier(int idx) const { 79 return _load_reference_barriers->at(idx); 80 } 81 82 void ShenandoahBarrierSetC2State::add_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) { 83 assert(!_load_reference_barriers->contains(n), "duplicate entry in barrier list"); 84 _load_reference_barriers->append(n); 85 } 86 87 void ShenandoahBarrierSetC2State::remove_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) { 88 if (_load_reference_barriers->contains(n)) { 89 _load_reference_barriers->remove(n); 90 } 91 } 92 93 Node* ShenandoahBarrierSetC2::shenandoah_iu_barrier(GraphKit* kit, Node* obj) const { 94 if (ShenandoahIUBarrier) { 95 return kit->gvn().transform(new ShenandoahIUBarrierNode(obj)); 96 } 97 return obj; 98 } 99 100 #define __ kit-> 101 102 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseValues* phase, Node* adr, 103 BasicType bt, uint adr_idx) const { 104 intptr_t offset = 0; 105 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 106 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 107 108 if (offset == Type::OffsetBot) { 109 return false; // cannot unalias unless there are precise offsets 110 } 111 112 if (alloc == nullptr) { 113 return false; // No allocation found 114 } 115 116 intptr_t size_in_bytes = type2aelembytes(bt); 117 118 Node* mem = __ memory(adr_idx); // start searching here... 119 120 for (int cnt = 0; cnt < 50; cnt++) { 121 122 if (mem->is_Store()) { 123 124 Node* st_adr = mem->in(MemNode::Address); 125 intptr_t st_offset = 0; 126 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 127 128 if (st_base == nullptr) { 129 break; // inscrutable pointer 130 } 131 132 // Break we have found a store with same base and offset as ours so break 133 if (st_base == base && st_offset == offset) { 134 break; 135 } 136 137 if (st_offset != offset && st_offset != Type::OffsetBot) { 138 const int MAX_STORE = BytesPerLong; 139 if (st_offset >= offset + size_in_bytes || 140 st_offset <= offset - MAX_STORE || 141 st_offset <= offset - mem->as_Store()->memory_size()) { 142 // Success: The offsets are provably independent. 143 // (You may ask, why not just test st_offset != offset and be done? 144 // The answer is that stores of different sizes can co-exist 145 // in the same sequence of RawMem effects. We sometimes initialize 146 // a whole 'tile' of array elements with a single jint or jlong.) 147 mem = mem->in(MemNode::Memory); 148 continue; // advance through independent store memory 149 } 150 } 151 152 if (st_base != base 153 && MemNode::detect_ptr_independence(base, alloc, st_base, 154 AllocateNode::Ideal_allocation(st_base, phase), 155 phase)) { 156 // Success: The bases are provably independent. 157 mem = mem->in(MemNode::Memory); 158 continue; // advance through independent store memory 159 } 160 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 161 162 InitializeNode* st_init = mem->in(0)->as_Initialize(); 163 AllocateNode* st_alloc = st_init->allocation(); 164 165 // Make sure that we are looking at the same allocation site. 166 // The alloc variable is guaranteed to not be null here from earlier check. 167 if (alloc == st_alloc) { 168 // Check that the initialization is storing null so that no previous store 169 // has been moved up and directly write a reference 170 Node* captured_store = st_init->find_captured_store(offset, 171 type2aelembytes(T_OBJECT), 172 phase); 173 if (captured_store == nullptr || captured_store == st_init->zero_memory()) { 174 return true; 175 } 176 } 177 } 178 179 // Unless there is an explicit 'continue', we must bail out here, 180 // because 'mem' is an inscrutable memory state (e.g., a call). 181 break; 182 } 183 184 return false; 185 } 186 187 #undef __ 188 #define __ ideal. 189 190 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit, 191 bool do_load, 192 Node* obj, 193 Node* adr, 194 uint alias_idx, 195 Node* val, 196 const TypeOopPtr* val_type, 197 Node* pre_val, 198 BasicType bt) const { 199 // Some sanity checks 200 // Note: val is unused in this routine. 201 202 if (do_load) { 203 // We need to generate the load of the previous value 204 assert(adr != nullptr, "where are loading from?"); 205 assert(pre_val == nullptr, "loaded already?"); 206 assert(val_type != nullptr, "need a type"); 207 208 if (ReduceInitialCardMarks 209 && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) { 210 return; 211 } 212 213 } else { 214 // In this case both val_type and alias_idx are unused. 215 assert(pre_val != nullptr, "must be loaded already"); 216 // Nothing to be done if pre_val is null. 217 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 218 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 219 } 220 assert(bt == T_OBJECT, "or we shouldn't be here"); 221 222 IdealKit ideal(kit, true); 223 224 Node* tls = __ thread(); // ThreadLocalStorage 225 226 Node* no_base = __ top(); 227 Node* zero = __ ConI(0); 228 Node* zeroX = __ ConX(0); 229 230 float likely = PROB_LIKELY(0.999); 231 float unlikely = PROB_UNLIKELY(0.999); 232 233 // Offsets into the thread 234 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()); 235 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()); 236 237 // Now the actual pointers into the thread 238 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 239 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 240 241 // Now some of the values 242 Node* marking; 243 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()))); 244 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw); 245 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING)); 246 assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape"); 247 248 // if (!marking) 249 __ if_then(marking, BoolTest::ne, zero, unlikely); { 250 BasicType index_bt = TypeX_X->basic_type(); 251 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading Shenandoah SATBMarkQueue::_index with wrong size."); 252 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 253 254 if (do_load) { 255 // load original value 256 // alias_idx correct?? 257 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 258 } 259 260 // if (pre_val != nullptr) 261 __ if_then(pre_val, BoolTest::ne, kit->null()); { 262 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 263 264 // is the queue for this thread full? 265 __ if_then(index, BoolTest::ne, zeroX, likely); { 266 267 // decrement the index 268 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 269 270 // Now get the buffer location we will log the previous value into and store it 271 Node *log_addr = __ AddP(no_base, buffer, next_index); 272 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 273 // update the index 274 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 275 276 } __ else_(); { 277 278 // logging buffer is full, call the runtime 279 const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type(); 280 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls); 281 } __ end_if(); // (!index) 282 } __ end_if(); // (pre_val != nullptr) 283 } __ end_if(); // (!marking) 284 285 // Final sync IdealKit and GraphKit. 286 kit->final_sync(ideal); 287 288 if (ShenandoahSATBBarrier && adr != nullptr) { 289 Node* c = kit->control(); 290 Node* call = c->in(1)->in(1)->in(1)->in(0); 291 assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected"); 292 call->add_req(adr); 293 } 294 } 295 296 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) { 297 return call->is_CallLeaf() && 298 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry); 299 } 300 301 bool ShenandoahBarrierSetC2::is_shenandoah_lrb_call(Node* call) { 302 if (!call->is_CallLeaf()) { 303 return false; 304 } 305 306 address entry_point = call->as_CallLeaf()->entry_point(); 307 return (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong)) || 308 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow)) || 309 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak)) || 310 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow)) || 311 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom)); 312 } 313 314 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseValues* phase, Node* n) { 315 if (n->Opcode() != Op_If) { 316 return false; 317 } 318 319 Node* bol = n->in(1); 320 assert(bol->is_Bool(), ""); 321 Node* cmpx = bol->in(1); 322 if (bol->as_Bool()->_test._test == BoolTest::ne && 323 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) && 324 is_shenandoah_state_load(cmpx->in(1)->in(1)) && 325 cmpx->in(1)->in(2)->is_Con() && 326 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) { 327 return true; 328 } 329 330 return false; 331 } 332 333 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) { 334 if (!n->is_Load()) return false; 335 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset()); 336 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal 337 && n->in(2)->in(3)->is_Con() 338 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset; 339 } 340 341 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit, 342 bool do_load, 343 Node* obj, 344 Node* adr, 345 uint alias_idx, 346 Node* val, 347 const TypeOopPtr* val_type, 348 Node* pre_val, 349 BasicType bt) const { 350 if (ShenandoahSATBBarrier) { 351 IdealKit ideal(kit); 352 kit->sync_kit(ideal); 353 354 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt); 355 356 ideal.sync_kit(kit); 357 kit->final_sync(ideal); 358 } 359 } 360 361 // Helper that guards and inserts a pre-barrier. 362 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset, 363 Node* pre_val, bool need_mem_bar) const { 364 // We could be accessing the referent field of a reference object. If so, when Shenandoah 365 // is enabled, we need to log the value in the referent field in an SATB buffer. 366 // This routine performs some compile time filters and generates suitable 367 // runtime filters that guard the pre-barrier code. 368 // Also add memory barrier for non volatile load from the referent field 369 // to prevent commoning of loads across safepoint. 370 371 // Some compile time checks. 372 373 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 374 const TypeX* otype = offset->find_intptr_t_type(); 375 if (otype != nullptr && otype->is_con() && 376 otype->get_con() != java_lang_ref_Reference::referent_offset()) { 377 // Constant offset but not the reference_offset so just return 378 return; 379 } 380 381 // We only need to generate the runtime guards for instances. 382 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 383 if (btype != nullptr) { 384 if (btype->isa_aryptr()) { 385 // Array type so nothing to do 386 return; 387 } 388 389 const TypeInstPtr* itype = btype->isa_instptr(); 390 if (itype != nullptr) { 391 // Can the klass of base_oop be statically determined to be 392 // _not_ a sub-class of Reference and _not_ Object? 393 ciKlass* klass = itype->instance_klass(); 394 if (klass->is_loaded() && 395 !klass->is_subtype_of(kit->env()->Reference_klass()) && 396 !kit->env()->Object_klass()->is_subtype_of(klass)) { 397 return; 398 } 399 } 400 } 401 402 // The compile time filters did not reject base_oop/offset so 403 // we need to generate the following runtime filters 404 // 405 // if (offset == java_lang_ref_Reference::_reference_offset) { 406 // if (instance_of(base, java.lang.ref.Reference)) { 407 // pre_barrier(_, pre_val, ...); 408 // } 409 // } 410 411 float likely = PROB_LIKELY( 0.999); 412 float unlikely = PROB_UNLIKELY(0.999); 413 414 IdealKit ideal(kit); 415 416 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset()); 417 418 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 419 // Update graphKit memory and control from IdealKit. 420 kit->sync_kit(ideal); 421 422 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass())); 423 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con); 424 425 // Update IdealKit memory and control from graphKit. 426 __ sync_kit(kit); 427 428 Node* one = __ ConI(1); 429 // is_instof == 0 if base_oop == nullptr 430 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 431 432 // Update graphKit from IdeakKit. 433 kit->sync_kit(ideal); 434 435 // Use the pre-barrier to record the value in the referent field 436 satb_write_barrier_pre(kit, false /* do_load */, 437 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */, 438 pre_val /* pre_val */, 439 T_OBJECT); 440 if (need_mem_bar) { 441 // Add memory barrier to prevent commoning reads from this field 442 // across safepoint since GC can change its value. 443 kit->insert_mem_bar(Op_MemBarCPUOrder); 444 } 445 // Update IdealKit from graphKit. 446 __ sync_kit(kit); 447 448 } __ end_if(); // _ref_type != ref_none 449 } __ end_if(); // offset == referent_offset 450 451 // Final sync IdealKit and GraphKit. 452 kit->final_sync(ideal); 453 } 454 455 Node* ShenandoahBarrierSetC2::byte_map_base_node(GraphKit* kit) const { 456 BarrierSet* bs = BarrierSet::barrier_set(); 457 ShenandoahBarrierSet* ctbs = barrier_set_cast<ShenandoahBarrierSet>(bs); 458 CardTable::CardValue* card_table_base = ctbs->card_table()->byte_map_base(); 459 if (card_table_base != nullptr) { 460 return kit->makecon(TypeRawPtr::make((address)card_table_base)); 461 } else { 462 return kit->null(); 463 } 464 } 465 466 void ShenandoahBarrierSetC2::post_barrier(GraphKit* kit, 467 Node* ctl, 468 Node* oop_store, 469 Node* obj, 470 Node* adr, 471 uint adr_idx, 472 Node* val, 473 BasicType bt, 474 bool use_precise) const { 475 assert(ShenandoahCardBarrier, "Did you mean to enable ShenandoahCardBarrier?"); 476 477 // No store check needed if we're storing a null. 478 if (val != nullptr && val->is_Con()) { 479 // must be either an oop or NULL 480 const Type* t = val->bottom_type(); 481 if (t == TypePtr::NULL_PTR || t == Type::TOP) 482 return; 483 } 484 485 if (ReduceInitialCardMarks && obj == kit->just_allocated_object(kit->control())) { 486 // We can skip marks on a freshly-allocated object in Eden. 487 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp. 488 // That routine informs GC to take appropriate compensating steps, 489 // upon a slow-path allocation, so as to make this card-mark 490 // elision safe. 491 return; 492 } 493 494 if (!use_precise) { 495 // All card marks for a (non-array) instance are in one place: 496 adr = obj; 497 } 498 // (Else it's an array (or unknown), and we want more precise card marks.) 499 assert(adr != nullptr, ""); 500 501 IdealKit ideal(kit, true); 502 503 // Convert the pointer to an int prior to doing math on it 504 Node* cast = __ CastPX(__ ctrl(), adr); 505 506 // Divide by card size 507 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift()) ); 508 509 // Combine card table base and card offset 510 Node* card_adr = __ AddP(__ top(), byte_map_base_node(kit), card_offset ); 511 512 // Get the alias_index for raw card-mark memory 513 int adr_type = Compile::AliasIdxRaw; 514 Node* zero = __ ConI(0); // Dirty card value 515 516 if (UseCondCardMark) { 517 // The classic GC reference write barrier is typically implemented 518 // as a store into the global card mark table. Unfortunately 519 // unconditional stores can result in false sharing and excessive 520 // coherence traffic as well as false transactional aborts. 521 // UseCondCardMark enables MP "polite" conditional card mark 522 // stores. In theory we could relax the load from ctrl() to 523 // no_ctrl, but that doesn't buy much latitude. 524 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, T_BYTE, adr_type); 525 __ if_then(card_val, BoolTest::ne, zero); 526 } 527 528 // Smash zero into card 529 __ store(__ ctrl(), card_adr, zero, T_BYTE, adr_type, MemNode::unordered); 530 531 if (UseCondCardMark) { 532 __ end_if(); 533 } 534 535 // Final sync IdealKit and GraphKit. 536 kit->final_sync(ideal); 537 } 538 539 #undef __ 540 541 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() { 542 const Type **fields = TypeTuple::fields(2); 543 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value 544 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread 545 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields); 546 547 // create result type (range) 548 fields = TypeTuple::fields(0); 549 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields); 550 551 return TypeFunc::make(domain, range); 552 } 553 554 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() { 555 const Type **fields = TypeTuple::fields(1); 556 fields[TypeFunc::Parms+0] = TypeOopPtr::NOTNULL; // src oop 557 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields); 558 559 // create result type (range) 560 fields = TypeTuple::fields(0); 561 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields); 562 563 return TypeFunc::make(domain, range); 564 } 565 566 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type() { 567 const Type **fields = TypeTuple::fields(2); 568 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM; // original field value 569 fields[TypeFunc::Parms+1] = TypeRawPtr::BOTTOM; // original load address 570 571 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields); 572 573 // create result type (range) 574 fields = TypeTuple::fields(1); 575 fields[TypeFunc::Parms+0] = TypeOopPtr::BOTTOM; 576 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields); 577 578 return TypeFunc::make(domain, range); 579 } 580 581 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const { 582 DecoratorSet decorators = access.decorators(); 583 584 const TypePtr* adr_type = access.addr().type(); 585 Node* adr = access.addr().node(); 586 587 if (!access.is_oop()) { 588 return BarrierSetC2::store_at_resolved(access, val); 589 } 590 591 if (access.is_parse_access()) { 592 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access); 593 GraphKit* kit = parse_access.kit(); 594 595 uint adr_idx = kit->C->get_alias_index(adr_type); 596 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 597 Node* value = val.node(); 598 value = shenandoah_iu_barrier(kit, value); 599 val.set_node(value); 600 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(), 601 static_cast<const TypeOopPtr*>(val.type()), nullptr /* pre_val */, access.type()); 602 603 Node* result = BarrierSetC2::store_at_resolved(access, val); 604 605 if (ShenandoahCardBarrier) { 606 const bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0; 607 const bool is_array = (decorators & IS_ARRAY) != 0; 608 const bool use_precise = is_array || anonymous; 609 post_barrier(kit, kit->control(), access.raw_access(), access.base(), 610 adr, adr_idx, val.node(), access.type(), use_precise); 611 } 612 return result; 613 } else { 614 assert(access.is_opt_access(), "only for optimization passes"); 615 assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code"); 616 C2OptAccess& opt_access = static_cast<C2OptAccess&>(access); 617 PhaseGVN& gvn = opt_access.gvn(); 618 619 if (ShenandoahIUBarrier) { 620 Node* enqueue = gvn.transform(new ShenandoahIUBarrierNode(val.node())); 621 val.set_node(enqueue); 622 } 623 return BarrierSetC2::store_at_resolved(access, val); 624 } 625 } 626 627 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const { 628 // 1: non-reference load, no additional barrier is needed 629 if (!access.is_oop()) { 630 return BarrierSetC2::load_at_resolved(access, val_type); 631 } 632 633 Node* load = BarrierSetC2::load_at_resolved(access, val_type); 634 DecoratorSet decorators = access.decorators(); 635 BasicType type = access.type(); 636 637 // 2: apply LRB if needed 638 if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) { 639 load = new ShenandoahLoadReferenceBarrierNode(nullptr, load, decorators); 640 if (access.is_parse_access()) { 641 load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load); 642 } else { 643 load = static_cast<C2OptAccess &>(access).gvn().transform(load); 644 } 645 } 646 647 // 3: apply keep-alive barrier for java.lang.ref.Reference if needed 648 if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) { 649 Node* top = Compile::current()->top(); 650 Node* adr = access.addr().node(); 651 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top; 652 Node* obj = access.base(); 653 654 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0; 655 bool on_weak_ref = (decorators & (ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF)) != 0; 656 bool keep_alive = (decorators & AS_NO_KEEPALIVE) == 0; 657 658 // If we are reading the value of the referent field of a Reference 659 // object (either by using Unsafe directly or through reflection) 660 // then, if SATB is enabled, we need to record the referent in an 661 // SATB log buffer using the pre-barrier mechanism. 662 // Also we need to add memory barrier to prevent commoning reads 663 // from this field across safepoint since GC can change its value. 664 if (!on_weak_ref || (unknown && (offset == top || obj == top)) || !keep_alive) { 665 return load; 666 } 667 668 assert(access.is_parse_access(), "entry not supported at optimization time"); 669 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access); 670 GraphKit* kit = parse_access.kit(); 671 bool mismatched = (decorators & C2_MISMATCHED) != 0; 672 bool is_unordered = (decorators & MO_UNORDERED) != 0; 673 bool in_native = (decorators & IN_NATIVE) != 0; 674 bool need_cpu_mem_bar = !is_unordered || mismatched || in_native; 675 676 if (on_weak_ref) { 677 // Use the pre-barrier to record the value in the referent field 678 satb_write_barrier_pre(kit, false /* do_load */, 679 nullptr /* obj */, nullptr /* adr */, max_juint /* alias_idx */, nullptr /* val */, nullptr /* val_type */, 680 load /* pre_val */, T_OBJECT); 681 // Add memory barrier to prevent commoning reads from this field 682 // across safepoint since GC can change its value. 683 kit->insert_mem_bar(Op_MemBarCPUOrder); 684 } else if (unknown) { 685 // We do not require a mem bar inside pre_barrier if need_mem_bar 686 // is set: the barriers would be emitted by us. 687 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar); 688 } 689 } 690 691 return load; 692 } 693 694 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val, 695 Node* new_val, const Type* value_type) const { 696 GraphKit* kit = access.kit(); 697 if (access.is_oop()) { 698 new_val = shenandoah_iu_barrier(kit, new_val); 699 shenandoah_write_barrier_pre(kit, false /* do_load */, 700 nullptr, nullptr, max_juint, nullptr, nullptr, 701 expected_val /* pre_val */, T_OBJECT); 702 703 MemNode::MemOrd mo = access.mem_node_mo(); 704 Node* mem = access.memory(); 705 Node* adr = access.addr().node(); 706 const TypePtr* adr_type = access.addr().type(); 707 Node* load_store = nullptr; 708 709 #ifdef _LP64 710 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 711 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop())); 712 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop())); 713 if (ShenandoahCASBarrier) { 714 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo)); 715 } else { 716 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo)); 717 } 718 } else 719 #endif 720 { 721 if (ShenandoahCASBarrier) { 722 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo)); 723 } else { 724 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo)); 725 } 726 } 727 728 access.set_raw_access(load_store); 729 pin_atomic_op(access); 730 731 #ifdef _LP64 732 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 733 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 734 } 735 #endif 736 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(nullptr, load_store, access.decorators())); 737 if (ShenandoahCardBarrier) { 738 post_barrier(kit, kit->control(), access.raw_access(), access.base(), 739 access.addr().node(), access.alias_idx(), new_val, T_OBJECT, true); 740 } 741 return load_store; 742 } 743 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type); 744 } 745 746 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val, 747 Node* new_val, const Type* value_type) const { 748 GraphKit* kit = access.kit(); 749 if (access.is_oop()) { 750 new_val = shenandoah_iu_barrier(kit, new_val); 751 shenandoah_write_barrier_pre(kit, false /* do_load */, 752 nullptr, nullptr, max_juint, nullptr, nullptr, 753 expected_val /* pre_val */, T_OBJECT); 754 DecoratorSet decorators = access.decorators(); 755 MemNode::MemOrd mo = access.mem_node_mo(); 756 Node* mem = access.memory(); 757 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0; 758 Node* load_store = nullptr; 759 Node* adr = access.addr().node(); 760 #ifdef _LP64 761 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 762 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop())); 763 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop())); 764 if (ShenandoahCASBarrier) { 765 if (is_weak_cas) { 766 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 767 } else { 768 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 769 } 770 } else { 771 if (is_weak_cas) { 772 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 773 } else { 774 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 775 } 776 } 777 } else 778 #endif 779 { 780 if (ShenandoahCASBarrier) { 781 if (is_weak_cas) { 782 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 783 } else { 784 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 785 } 786 } else { 787 if (is_weak_cas) { 788 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 789 } else { 790 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 791 } 792 } 793 } 794 access.set_raw_access(load_store); 795 pin_atomic_op(access); 796 if (ShenandoahCardBarrier) { 797 post_barrier(kit, kit->control(), access.raw_access(), access.base(), 798 access.addr().node(), access.alias_idx(), new_val, T_OBJECT, true); 799 } 800 return load_store; 801 } 802 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type); 803 } 804 805 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const { 806 GraphKit* kit = access.kit(); 807 if (access.is_oop()) { 808 val = shenandoah_iu_barrier(kit, val); 809 } 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 // Support for GC barriers emitted during parsing 830 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const { 831 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier || node->Opcode() == Op_ShenandoahIUBarrier) return true; 832 if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) { 833 return false; 834 } 835 CallLeafNode *call = node->as_CallLeaf(); 836 if (call->_name == nullptr) { 837 return false; 838 } 839 840 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 || 841 strcmp(call->_name, "shenandoah_cas_obj") == 0 || 842 strcmp(call->_name, "shenandoah_wb_pre") == 0; 843 } 844 845 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const { 846 if (c == nullptr) { 847 return c; 848 } 849 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) { 850 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn); 851 } 852 if (c->Opcode() == Op_ShenandoahIUBarrier) { 853 c = c->in(1); 854 } 855 return c; 856 } 857 858 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const { 859 return !ShenandoahBarrierC2Support::expand(C, igvn); 860 } 861 862 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const { 863 if (mode == LoopOptsShenandoahExpand) { 864 assert(UseShenandoahGC, "only for shenandoah"); 865 ShenandoahBarrierC2Support::pin_and_expand(phase); 866 return true; 867 } else if (mode == LoopOptsShenandoahPostExpand) { 868 assert(UseShenandoahGC, "only for shenandoah"); 869 visited.clear(); 870 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase); 871 return true; 872 } 873 return false; 874 } 875 876 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, bool is_clone_instance, ArrayCopyPhase phase) const { 877 bool is_oop = is_reference_type(type); 878 if (!is_oop) { 879 return false; 880 } 881 if (ShenandoahSATBBarrier && tightly_coupled_alloc) { 882 if (phase == Optimization) { 883 return false; 884 } 885 return !is_clone; 886 } 887 if (phase == Optimization) { 888 return !ShenandoahIUBarrier; 889 } 890 return true; 891 } 892 893 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) { 894 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr(); 895 if (src_type->isa_instptr() != nullptr) { 896 ciInstanceKlass* ik = src_type->is_instptr()->instance_klass(); 897 if ((src_type->klass_is_exact() || !ik->has_subklass()) && !ik->has_injected_fields()) { 898 if (ik->has_object_fields()) { 899 return true; 900 } else { 901 if (!src_type->klass_is_exact()) { 902 Compile::current()->dependencies()->assert_leaf_type(ik); 903 } 904 } 905 } else { 906 return true; 907 } 908 } else if (src_type->isa_aryptr()) { 909 BasicType src_elem = src_type->isa_aryptr()->elem()->array_element_basic_type(); 910 if (is_reference_type(src_elem, true)) { 911 return true; 912 } 913 } else { 914 return true; 915 } 916 return false; 917 } 918 919 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const { 920 Node* ctrl = ac->in(TypeFunc::Control); 921 Node* mem = ac->in(TypeFunc::Memory); 922 Node* src_base = ac->in(ArrayCopyNode::Src); 923 Node* src_offset = ac->in(ArrayCopyNode::SrcPos); 924 Node* dest_base = ac->in(ArrayCopyNode::Dest); 925 Node* dest_offset = ac->in(ArrayCopyNode::DestPos); 926 Node* length = ac->in(ArrayCopyNode::Length); 927 928 Node* src = phase->basic_plus_adr(src_base, src_offset); 929 Node* dest = phase->basic_plus_adr(dest_base, dest_offset); 930 931 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) { 932 // Check if heap is has forwarded objects. If it does, we need to call into the special 933 // routine that would fix up source references before we can continue. 934 935 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT }; 936 Node* region = new RegionNode(PATH_LIMIT); 937 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM); 938 939 Node* thread = phase->transform_later(new ThreadLocalNode()); 940 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())); 941 Node* gc_state_addr = phase->transform_later(new AddPNode(phase->C->top(), thread, offset)); 942 943 uint gc_state_idx = Compile::AliasIdxRaw; 944 const TypePtr* gc_state_adr_type = nullptr; // debug-mode-only argument 945 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx)); 946 947 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered)); 948 int flags = ShenandoahHeap::HAS_FORWARDED; 949 if (ShenandoahIUBarrier) { 950 flags |= ShenandoahHeap::MARKING; 951 } 952 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(flags))); 953 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT))); 954 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne)); 955 956 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If(); 957 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff)); 958 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff)); 959 960 // Heap is stable, no need to do anything additional 961 region->init_req(_heap_stable, stable_ctrl); 962 mem_phi->init_req(_heap_stable, mem); 963 964 // Heap is unstable, call into clone barrier stub 965 Node* call = phase->make_leaf_call(unstable_ctrl, mem, 966 ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(), 967 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier), 968 "shenandoah_clone", 969 TypeRawPtr::BOTTOM, 970 src_base); 971 call = phase->transform_later(call); 972 973 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control)); 974 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory)); 975 region->init_req(_heap_unstable, ctrl); 976 mem_phi->init_req(_heap_unstable, mem); 977 978 // Wire up the actual arraycopy stub now 979 ctrl = phase->transform_later(region); 980 mem = phase->transform_later(mem_phi); 981 982 const char* name = "arraycopy"; 983 call = phase->make_leaf_call(ctrl, mem, 984 OptoRuntime::fast_arraycopy_Type(), 985 phase->basictype2arraycopy(T_LONG, nullptr, nullptr, true, name, true), 986 name, TypeRawPtr::BOTTOM, 987 src, dest, length 988 LP64_ONLY(COMMA phase->top())); 989 call = phase->transform_later(call); 990 991 // Hook up the whole thing into the graph 992 phase->igvn().replace_node(ac, call); 993 } else { 994 BarrierSetC2::clone_at_expansion(phase, ac); 995 } 996 } 997 998 999 // Support for macro expanded GC barriers 1000 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const { 1001 if (node->Opcode() == Op_ShenandoahIUBarrier) { 1002 state()->add_iu_barrier((ShenandoahIUBarrierNode*) node); 1003 } 1004 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) { 1005 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node); 1006 } 1007 } 1008 1009 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const { 1010 if (node->Opcode() == Op_ShenandoahIUBarrier) { 1011 state()->remove_iu_barrier((ShenandoahIUBarrierNode*) node); 1012 } 1013 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) { 1014 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node); 1015 } 1016 } 1017 1018 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const { 1019 if (is_shenandoah_wb_pre_call(node)) { 1020 shenandoah_eliminate_wb_pre(node, ¯o->igvn()); 1021 } 1022 if (ShenandoahCardBarrier && node->Opcode() == Op_CastP2X) { 1023 Node* shift = node->unique_out(); 1024 Node* addp = shift->unique_out(); 1025 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) { 1026 Node* mem = addp->last_out(j); 1027 if (UseCondCardMark && mem->is_Load()) { 1028 assert(mem->Opcode() == Op_LoadB, "unexpected code shape"); 1029 // The load is checking if the card has been written so 1030 // replace it with zero to fold the test. 1031 macro->replace_node(mem, macro->intcon(0)); 1032 continue; 1033 } 1034 assert(mem->is_Store(), "store required"); 1035 macro->replace_node(mem, mem->in(MemNode::Memory)); 1036 } 1037 } 1038 } 1039 1040 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const { 1041 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), ""); 1042 Node* c = call->as_Call()->proj_out(TypeFunc::Control); 1043 c = c->unique_ctrl_out(); 1044 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 1045 c = c->unique_ctrl_out(); 1046 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 1047 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0); 1048 assert(iff->is_If(), "expect test"); 1049 if (!is_shenandoah_marking_if(igvn, iff)) { 1050 c = c->unique_ctrl_out(); 1051 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 1052 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0); 1053 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test"); 1054 } 1055 Node* cmpx = iff->in(1)->in(1); 1056 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ)); 1057 igvn->rehash_node_delayed(call); 1058 call->del_req(call->req()-1); 1059 } 1060 1061 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const { 1062 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) { 1063 igvn->add_users_to_worklist(node); 1064 } 1065 } 1066 1067 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const { 1068 for (uint i = 0; i < useful.size(); i++) { 1069 Node* n = useful.at(i); 1070 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) { 1071 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1072 C->record_for_igvn(n->fast_out(i)); 1073 } 1074 } 1075 } 1076 for (int i = state()->iu_barriers_count() - 1; i >= 0; i--) { 1077 ShenandoahIUBarrierNode* n = state()->iu_barrier(i); 1078 if (!useful.member(n)) { 1079 state()->remove_iu_barrier(n); 1080 } 1081 } 1082 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) { 1083 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i); 1084 if (!useful.member(n)) { 1085 state()->remove_load_reference_barrier(n); 1086 } 1087 } 1088 } 1089 1090 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const { 1091 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena); 1092 } 1093 1094 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const { 1095 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state()); 1096 } 1097 1098 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be 1099 // expanded later, then now is the time to do so. 1100 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; } 1101 1102 #ifdef ASSERT 1103 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const { 1104 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) { 1105 ShenandoahBarrierC2Support::verify(Compile::current()->root()); 1106 } else if (phase == BarrierSetC2::BeforeCodeGen) { 1107 // Verify Shenandoah pre-barriers 1108 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()); 1109 1110 Unique_Node_List visited; 1111 Node_List worklist; 1112 // We're going to walk control flow backwards starting from the Root 1113 worklist.push(compile->root()); 1114 while (worklist.size() > 0) { 1115 Node *x = worklist.pop(); 1116 if (x == nullptr || x == compile->top()) continue; 1117 if (visited.member(x)) { 1118 continue; 1119 } else { 1120 visited.push(x); 1121 } 1122 1123 if (x->is_Region()) { 1124 for (uint i = 1; i < x->req(); i++) { 1125 worklist.push(x->in(i)); 1126 } 1127 } else { 1128 worklist.push(x->in(0)); 1129 // We are looking for the pattern: 1130 // /->ThreadLocal 1131 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 1132 // \->ConI(0) 1133 // We want to verify that the If and the LoadB have the same control 1134 // See GraphKit::g1_write_barrier_pre() 1135 if (x->is_If()) { 1136 IfNode *iff = x->as_If(); 1137 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 1138 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 1139 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 1140 && cmp->in(1)->is_Load()) { 1141 LoadNode *load = cmp->in(1)->as_Load(); 1142 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 1143 && load->in(2)->in(3)->is_Con() 1144 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 1145 1146 Node *if_ctrl = iff->in(0); 1147 Node *load_ctrl = load->in(0); 1148 1149 if (if_ctrl != load_ctrl) { 1150 // Skip possible CProj->NeverBranch in infinite loops 1151 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 1152 && if_ctrl->in(0)->is_NeverBranch()) { 1153 if_ctrl = if_ctrl->in(0)->in(0); 1154 } 1155 } 1156 assert(load_ctrl != nullptr && if_ctrl == load_ctrl, "controls must match"); 1157 } 1158 } 1159 } 1160 } 1161 } 1162 } 1163 } 1164 } 1165 #endif 1166 1167 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const { 1168 if (is_shenandoah_wb_pre_call(n)) { 1169 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt(); 1170 if (n->req() > cnt) { 1171 Node* addp = n->in(cnt); 1172 if (has_only_shenandoah_wb_pre_uses(addp)) { 1173 n->del_req(cnt); 1174 if (can_reshape) { 1175 phase->is_IterGVN()->_worklist.push(addp); 1176 } 1177 return n; 1178 } 1179 } 1180 } 1181 if (n->Opcode() == Op_CmpP) { 1182 Node* in1 = n->in(1); 1183 Node* in2 = n->in(2); 1184 1185 // If one input is null, then step over the strong LRB barriers on the other input 1186 if (in1->bottom_type() == TypePtr::NULL_PTR && 1187 !((in2->Opcode() == Op_ShenandoahLoadReferenceBarrier) && 1188 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in2)->decorators()))) { 1189 in2 = step_over_gc_barrier(in2); 1190 } 1191 if (in2->bottom_type() == TypePtr::NULL_PTR && 1192 !((in1->Opcode() == Op_ShenandoahLoadReferenceBarrier) && 1193 !ShenandoahBarrierSet::is_strong_access(((ShenandoahLoadReferenceBarrierNode*)in1)->decorators()))) { 1194 in1 = step_over_gc_barrier(in1); 1195 } 1196 1197 if (in1 != n->in(1)) { 1198 n->set_req_X(1, in1, phase); 1199 assert(in2 == n->in(2), "only one change"); 1200 return n; 1201 } 1202 if (in2 != n->in(2)) { 1203 n->set_req_X(2, in2, phase); 1204 return n; 1205 } 1206 } else if (can_reshape && 1207 n->Opcode() == Op_If && 1208 ShenandoahBarrierC2Support::is_heap_stable_test(n) && 1209 n->in(0) != nullptr && 1210 n->outcnt() == 2) { 1211 Node* dom = n->in(0); 1212 Node* prev_dom = n; 1213 int op = n->Opcode(); 1214 int dist = 16; 1215 // Search up the dominator tree for another heap stable test 1216 while (dom->Opcode() != op || // Not same opcode? 1217 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1? 1218 prev_dom->in(0) != dom) { // One path of test does not dominate? 1219 if (dist < 0) return nullptr; 1220 1221 dist--; 1222 prev_dom = dom; 1223 dom = IfNode::up_one_dom(dom); 1224 if (!dom) return nullptr; 1225 } 1226 1227 // Check that we did not follow a loop back to ourselves 1228 if (n == dom) { 1229 return nullptr; 1230 } 1231 1232 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN(), false); 1233 } 1234 1235 return nullptr; 1236 } 1237 1238 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) { 1239 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1240 Node* u = n->fast_out(i); 1241 if (!is_shenandoah_wb_pre_call(u)) { 1242 return false; 1243 } 1244 } 1245 return n->outcnt() > 0; 1246 } 1247 1248 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode, Unique_Node_List& dead_nodes) const { 1249 switch (opcode) { 1250 case Op_CallLeaf: 1251 case Op_CallLeafNoFP: { 1252 assert (n->is_Call(), ""); 1253 CallNode *call = n->as_Call(); 1254 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) { 1255 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt(); 1256 if (call->req() > cnt) { 1257 assert(call->req() == cnt + 1, "only one extra input"); 1258 Node *addp = call->in(cnt); 1259 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?"); 1260 call->del_req(cnt); 1261 } 1262 } 1263 return false; 1264 } 1265 case Op_ShenandoahCompareAndSwapP: 1266 case Op_ShenandoahCompareAndSwapN: 1267 case Op_ShenandoahWeakCompareAndSwapN: 1268 case Op_ShenandoahWeakCompareAndSwapP: 1269 case Op_ShenandoahCompareAndExchangeP: 1270 case Op_ShenandoahCompareAndExchangeN: 1271 return true; 1272 case Op_ShenandoahLoadReferenceBarrier: 1273 assert(false, "should have been expanded already"); 1274 return true; 1275 default: 1276 return false; 1277 } 1278 } 1279 1280 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const { 1281 switch (opcode) { 1282 case Op_ShenandoahCompareAndExchangeP: 1283 case Op_ShenandoahCompareAndExchangeN: 1284 conn_graph->add_objload_to_connection_graph(n, delayed_worklist); 1285 // fallthrough 1286 case Op_ShenandoahWeakCompareAndSwapP: 1287 case Op_ShenandoahWeakCompareAndSwapN: 1288 case Op_ShenandoahCompareAndSwapP: 1289 case Op_ShenandoahCompareAndSwapN: 1290 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist); 1291 return true; 1292 case Op_StoreP: { 1293 Node* adr = n->in(MemNode::Address); 1294 const Type* adr_type = gvn->type(adr); 1295 // Pointer stores in Shenandoah barriers looks like unsafe access. 1296 // Ignore such stores to be able scalar replace non-escaping 1297 // allocations. 1298 if (adr_type->isa_rawptr() && adr->is_AddP()) { 1299 Node* base = conn_graph->get_addp_base(adr); 1300 if (base->Opcode() == Op_LoadP && 1301 base->in(MemNode::Address)->is_AddP()) { 1302 adr = base->in(MemNode::Address); 1303 Node* tls = conn_graph->get_addp_base(adr); 1304 if (tls->Opcode() == Op_ThreadLocal) { 1305 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 1306 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()); 1307 if (offs == buf_offset) { 1308 return true; // Pre barrier previous oop value store. 1309 } 1310 } 1311 } 1312 } 1313 return false; 1314 } 1315 case Op_ShenandoahIUBarrier: 1316 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist); 1317 break; 1318 case Op_ShenandoahLoadReferenceBarrier: 1319 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist); 1320 return true; 1321 default: 1322 // Nothing 1323 break; 1324 } 1325 return false; 1326 } 1327 1328 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const { 1329 switch (opcode) { 1330 case Op_ShenandoahCompareAndExchangeP: 1331 case Op_ShenandoahCompareAndExchangeN: { 1332 Node *adr = n->in(MemNode::Address); 1333 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, nullptr); 1334 // fallthrough 1335 } 1336 case Op_ShenandoahCompareAndSwapP: 1337 case Op_ShenandoahCompareAndSwapN: 1338 case Op_ShenandoahWeakCompareAndSwapP: 1339 case Op_ShenandoahWeakCompareAndSwapN: 1340 return conn_graph->add_final_edges_unsafe_access(n, opcode); 1341 case Op_ShenandoahIUBarrier: 1342 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), nullptr); 1343 return true; 1344 case Op_ShenandoahLoadReferenceBarrier: 1345 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), nullptr); 1346 return true; 1347 default: 1348 // Nothing 1349 break; 1350 } 1351 return false; 1352 } 1353 1354 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const { 1355 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) || 1356 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN); 1357 1358 } 1359 1360 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const { 1361 switch (opcode) { 1362 case Op_ShenandoahCompareAndExchangeP: 1363 case Op_ShenandoahCompareAndExchangeN: 1364 case Op_ShenandoahWeakCompareAndSwapP: 1365 case Op_ShenandoahWeakCompareAndSwapN: 1366 case Op_ShenandoahCompareAndSwapP: 1367 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree 1368 Node* newval = n->in(MemNode::ValueIn); 1369 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn); 1370 Node* pair = new BinaryNode(oldval, newval); 1371 n->set_req(MemNode::ValueIn,pair); 1372 n->del_req(LoadStoreConditionalNode::ExpectedIn); 1373 return true; 1374 } 1375 default: 1376 break; 1377 } 1378 return false; 1379 } 1380 1381 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const { 1382 return xop == Op_ShenandoahCompareAndExchangeP || 1383 xop == Op_ShenandoahCompareAndExchangeN || 1384 xop == Op_ShenandoahWeakCompareAndSwapP || 1385 xop == Op_ShenandoahWeakCompareAndSwapN || 1386 xop == Op_ShenandoahCompareAndSwapN || 1387 xop == Op_ShenandoahCompareAndSwapP; 1388 }