1 /* 2 * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/shared/barrierSet.hpp" 27 #include "gc/shenandoah/shenandoahBarrierSet.hpp" 28 #include "gc/shenandoah/shenandoahForwarding.hpp" 29 #include "gc/shenandoah/shenandoahHeap.hpp" 30 #include "gc/shenandoah/shenandoahHeuristics.hpp" 31 #include "gc/shenandoah/shenandoahRuntime.hpp" 32 #include "gc/shenandoah/shenandoahThreadLocalData.hpp" 33 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp" 34 #include "gc/shenandoah/c2/shenandoahSupport.hpp" 35 #include "opto/arraycopynode.hpp" 36 #include "opto/escape.hpp" 37 #include "opto/graphKit.hpp" 38 #include "opto/idealKit.hpp" 39 #include "opto/macro.hpp" 40 #include "opto/movenode.hpp" 41 #include "opto/narrowptrnode.hpp" 42 #include "opto/rootnode.hpp" 43 #include "opto/runtime.hpp" 44 45 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() { 46 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2()); 47 } 48 49 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena) 50 : _enqueue_barriers(new (comp_arena) GrowableArray<ShenandoahEnqueueBarrierNode*>(comp_arena, 8, 0, NULL)), 51 _load_reference_barriers(new (comp_arena) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena, 8, 0, NULL)) { 52 } 53 54 int ShenandoahBarrierSetC2State::enqueue_barriers_count() const { 55 return _enqueue_barriers->length(); 56 } 57 58 ShenandoahEnqueueBarrierNode* ShenandoahBarrierSetC2State::enqueue_barrier(int idx) const { 59 return _enqueue_barriers->at(idx); 60 } 61 62 void ShenandoahBarrierSetC2State::add_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) { 63 assert(!_enqueue_barriers->contains(n), "duplicate entry in barrier list"); 64 _enqueue_barriers->append(n); 65 } 66 67 void ShenandoahBarrierSetC2State::remove_enqueue_barrier(ShenandoahEnqueueBarrierNode * n) { 68 if (_enqueue_barriers->contains(n)) { 69 _enqueue_barriers->remove(n); 70 } 71 } 72 73 int ShenandoahBarrierSetC2State::load_reference_barriers_count() const { 74 return _load_reference_barriers->length(); 75 } 76 77 ShenandoahLoadReferenceBarrierNode* ShenandoahBarrierSetC2State::load_reference_barrier(int idx) const { 78 return _load_reference_barriers->at(idx); 79 } 80 81 void ShenandoahBarrierSetC2State::add_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) { 82 assert(!_load_reference_barriers->contains(n), "duplicate entry in barrier list"); 83 _load_reference_barriers->append(n); 84 } 85 86 void ShenandoahBarrierSetC2State::remove_load_reference_barrier(ShenandoahLoadReferenceBarrierNode * n) { 87 if (_load_reference_barriers->contains(n)) { 88 _load_reference_barriers->remove(n); 89 } 90 } 91 92 Node* ShenandoahBarrierSetC2::shenandoah_storeval_barrier(GraphKit* kit, Node* obj) const { 93 if (ShenandoahStoreValEnqueueBarrier) { 94 obj = shenandoah_enqueue_barrier(kit, obj); 95 } 96 return obj; 97 } 98 99 #define __ kit-> 100 101 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseTransform* phase, Node* adr, 102 BasicType bt, uint adr_idx) const { 103 intptr_t offset = 0; 104 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 105 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 106 107 if (offset == Type::OffsetBot) { 108 return false; // cannot unalias unless there are precise offsets 109 } 110 111 if (alloc == NULL) { 112 return false; // No allocation found 113 } 114 115 intptr_t size_in_bytes = type2aelembytes(bt); 116 117 Node* mem = __ memory(adr_idx); // start searching here... 118 119 for (int cnt = 0; cnt < 50; cnt++) { 120 121 if (mem->is_Store()) { 122 123 Node* st_adr = mem->in(MemNode::Address); 124 intptr_t st_offset = 0; 125 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 126 127 if (st_base == NULL) { 128 break; // inscrutable pointer 129 } 130 131 // Break we have found a store with same base and offset as ours so break 132 if (st_base == base && st_offset == offset) { 133 break; 134 } 135 136 if (st_offset != offset && st_offset != Type::OffsetBot) { 137 const int MAX_STORE = BytesPerLong; 138 if (st_offset >= offset + size_in_bytes || 139 st_offset <= offset - MAX_STORE || 140 st_offset <= offset - mem->as_Store()->memory_size()) { 141 // Success: The offsets are provably independent. 142 // (You may ask, why not just test st_offset != offset and be done? 143 // The answer is that stores of different sizes can co-exist 144 // in the same sequence of RawMem effects. We sometimes initialize 145 // a whole 'tile' of array elements with a single jint or jlong.) 146 mem = mem->in(MemNode::Memory); 147 continue; // advance through independent store memory 148 } 149 } 150 151 if (st_base != base 152 && MemNode::detect_ptr_independence(base, alloc, st_base, 153 AllocateNode::Ideal_allocation(st_base, phase), 154 phase)) { 155 // Success: The bases are provably independent. 156 mem = mem->in(MemNode::Memory); 157 continue; // advance through independent store memory 158 } 159 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 160 161 InitializeNode* st_init = mem->in(0)->as_Initialize(); 162 AllocateNode* st_alloc = st_init->allocation(); 163 164 // Make sure that we are looking at the same allocation site. 165 // The alloc variable is guaranteed to not be null here from earlier check. 166 if (alloc == st_alloc) { 167 // Check that the initialization is storing NULL so that no previous store 168 // has been moved up and directly write a reference 169 Node* captured_store = st_init->find_captured_store(offset, 170 type2aelembytes(T_OBJECT), 171 phase); 172 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 173 return true; 174 } 175 } 176 } 177 178 // Unless there is an explicit 'continue', we must bail out here, 179 // because 'mem' is an inscrutable memory state (e.g., a call). 180 break; 181 } 182 183 return false; 184 } 185 186 #undef __ 187 #define __ ideal. 188 189 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit, 190 bool do_load, 191 Node* obj, 192 Node* adr, 193 uint alias_idx, 194 Node* val, 195 const TypeOopPtr* val_type, 196 Node* pre_val, 197 BasicType bt) const { 198 // Some sanity checks 199 // Note: val is unused in this routine. 200 201 if (do_load) { 202 // We need to generate the load of the previous value 203 assert(obj != NULL, "must have a base"); 204 assert(adr != NULL, "where are loading from?"); 205 assert(pre_val == NULL, "loaded already?"); 206 assert(val_type != NULL, "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 != NULL, "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 G1 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 != NULL) 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 != NULL) 283 } __ end_if(); // (!marking) 284 285 // Final sync IdealKit and GraphKit. 286 kit->final_sync(ideal); 287 288 if (ShenandoahSATBBarrier && adr != NULL) { 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)) || 308 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_narrow)) || 309 (entry_point == CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_native)); 310 } 311 312 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseTransform *phase, Node* n) { 313 if (n->Opcode() != Op_If) { 314 return false; 315 } 316 317 Node* bol = n->in(1); 318 assert(bol->is_Bool(), ""); 319 Node* cmpx = bol->in(1); 320 if (bol->as_Bool()->_test._test == BoolTest::ne && 321 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) && 322 is_shenandoah_state_load(cmpx->in(1)->in(1)) && 323 cmpx->in(1)->in(2)->is_Con() && 324 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) { 325 return true; 326 } 327 328 return false; 329 } 330 331 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) { 332 if (!n->is_Load()) return false; 333 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset()); 334 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal 335 && n->in(2)->in(3)->is_Con() 336 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset; 337 } 338 339 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit, 340 bool do_load, 341 Node* obj, 342 Node* adr, 343 uint alias_idx, 344 Node* val, 345 const TypeOopPtr* val_type, 346 Node* pre_val, 347 BasicType bt) const { 348 if (ShenandoahSATBBarrier) { 349 IdealKit ideal(kit); 350 kit->sync_kit(ideal); 351 352 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt); 353 354 ideal.sync_kit(kit); 355 kit->final_sync(ideal); 356 } 357 } 358 359 Node* ShenandoahBarrierSetC2::shenandoah_enqueue_barrier(GraphKit* kit, Node* pre_val) const { 360 return kit->gvn().transform(new ShenandoahEnqueueBarrierNode(pre_val)); 361 } 362 363 // Helper that guards and inserts a pre-barrier. 364 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset, 365 Node* pre_val, bool need_mem_bar) const { 366 // We could be accessing the referent field of a reference object. If so, when G1 367 // is enabled, we need to log the value in the referent field in an SATB buffer. 368 // This routine performs some compile time filters and generates suitable 369 // runtime filters that guard the pre-barrier code. 370 // Also add memory barrier for non volatile load from the referent field 371 // to prevent commoning of loads across safepoint. 372 373 // Some compile time checks. 374 375 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 376 const TypeX* otype = offset->find_intptr_t_type(); 377 if (otype != NULL && otype->is_con() && 378 otype->get_con() != java_lang_ref_Reference::referent_offset) { 379 // Constant offset but not the reference_offset so just return 380 return; 381 } 382 383 // We only need to generate the runtime guards for instances. 384 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 385 if (btype != NULL) { 386 if (btype->isa_aryptr()) { 387 // Array type so nothing to do 388 return; 389 } 390 391 const TypeInstPtr* itype = btype->isa_instptr(); 392 if (itype != NULL) { 393 // Can the klass of base_oop be statically determined to be 394 // _not_ a sub-class of Reference and _not_ Object? 395 ciKlass* klass = itype->klass(); 396 if ( klass->is_loaded() && 397 !klass->is_subtype_of(kit->env()->Reference_klass()) && 398 !kit->env()->Object_klass()->is_subtype_of(klass)) { 399 return; 400 } 401 } 402 } 403 404 // The compile time filters did not reject base_oop/offset so 405 // we need to generate the following runtime filters 406 // 407 // if (offset == java_lang_ref_Reference::_reference_offset) { 408 // if (instance_of(base, java.lang.ref.Reference)) { 409 // pre_barrier(_, pre_val, ...); 410 // } 411 // } 412 413 float likely = PROB_LIKELY( 0.999); 414 float unlikely = PROB_UNLIKELY(0.999); 415 416 IdealKit ideal(kit); 417 418 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 419 420 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 421 // Update graphKit memory and control from IdealKit. 422 kit->sync_kit(ideal); 423 424 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass())); 425 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con); 426 427 // Update IdealKit memory and control from graphKit. 428 __ sync_kit(kit); 429 430 Node* one = __ ConI(1); 431 // is_instof == 0 if base_oop == NULL 432 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 433 434 // Update graphKit from IdeakKit. 435 kit->sync_kit(ideal); 436 437 // Use the pre-barrier to record the value in the referent field 438 satb_write_barrier_pre(kit, false /* do_load */, 439 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 440 pre_val /* pre_val */, 441 T_OBJECT); 442 if (need_mem_bar) { 443 // Add memory barrier to prevent commoning reads from this field 444 // across safepoint since GC can change its value. 445 kit->insert_mem_bar(Op_MemBarCPUOrder); 446 } 447 // Update IdealKit from graphKit. 448 __ sync_kit(kit); 449 450 } __ end_if(); // _ref_type != ref_none 451 } __ end_if(); // offset == referent_offset 452 453 // Final sync IdealKit and GraphKit. 454 kit->final_sync(ideal); 455 } 456 457 #undef __ 458 459 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() { 460 const Type **fields = TypeTuple::fields(2); 461 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value 462 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread 463 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields); 464 465 // create result type (range) 466 fields = TypeTuple::fields(0); 467 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields); 468 469 return TypeFunc::make(domain, range); 470 } 471 472 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() { 473 const Type **fields = TypeTuple::fields(1); 474 fields[TypeFunc::Parms+0] = TypeOopPtr::NOTNULL; // src oop 475 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields); 476 477 // create result type (range) 478 fields = TypeTuple::fields(0); 479 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields); 480 481 return TypeFunc::make(domain, range); 482 } 483 484 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type() { 485 const Type **fields = TypeTuple::fields(2); 486 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value 487 fields[TypeFunc::Parms+1] = TypeRawPtr::BOTTOM; // original load address 488 489 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields); 490 491 // create result type (range) 492 fields = TypeTuple::fields(1); 493 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; 494 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields); 495 496 return TypeFunc::make(domain, range); 497 } 498 499 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const { 500 DecoratorSet decorators = access.decorators(); 501 502 const TypePtr* adr_type = access.addr().type(); 503 Node* adr = access.addr().node(); 504 505 bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0; 506 bool on_heap = (decorators & IN_HEAP) != 0; 507 508 if (!access.is_oop() || (!on_heap && !anonymous)) { 509 return BarrierSetC2::store_at_resolved(access, val); 510 } 511 512 if (access.is_parse_access()) { 513 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access); 514 GraphKit* kit = parse_access.kit(); 515 516 uint adr_idx = kit->C->get_alias_index(adr_type); 517 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 518 Node* value = val.node(); 519 value = shenandoah_storeval_barrier(kit, value); 520 val.set_node(value); 521 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(), 522 static_cast<const TypeOopPtr*>(val.type()), NULL /* pre_val */, access.type()); 523 } else { 524 assert(access.is_opt_access(), "only for optimization passes"); 525 assert(((decorators & C2_TIGHTLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code"); 526 C2OptAccess& opt_access = static_cast<C2OptAccess&>(access); 527 PhaseGVN& gvn = opt_access.gvn(); 528 MergeMemNode* mm = opt_access.mem(); 529 530 if (ShenandoahStoreValEnqueueBarrier) { 531 Node* enqueue = gvn.transform(new ShenandoahEnqueueBarrierNode(val.node())); 532 val.set_node(enqueue); 533 } 534 } 535 return BarrierSetC2::store_at_resolved(access, val); 536 } 537 538 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const { 539 // 1: non-reference load, no additional barrier is needed 540 if (!access.is_oop()) { 541 return BarrierSetC2::load_at_resolved(access, val_type);; 542 } 543 544 Node* load = BarrierSetC2::load_at_resolved(access, val_type); 545 DecoratorSet decorators = access.decorators(); 546 BasicType type = access.type(); 547 548 // 2: apply LRB if needed 549 if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) { 550 load = new ShenandoahLoadReferenceBarrierNode(NULL, 551 load, 552 ShenandoahBarrierSet::use_load_reference_barrier_native(decorators, type)); 553 if (access.is_parse_access()) { 554 load = static_cast<C2ParseAccess &>(access).kit()->gvn().transform(load); 555 } else { 556 load = static_cast<C2OptAccess &>(access).gvn().transform(load); 557 } 558 } 559 560 // 3: apply keep-alive barrier if needed 561 if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) { 562 Node* top = Compile::current()->top(); 563 Node* adr = access.addr().node(); 564 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top; 565 Node* obj = access.base(); 566 567 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0; 568 bool on_weak_ref = (decorators & (ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF)) != 0; 569 bool is_traversal_mode = ShenandoahHeap::heap()->is_traversal_mode(); 570 bool keep_alive = (decorators & AS_NO_KEEPALIVE) == 0 || is_traversal_mode; 571 572 // If we are reading the value of the referent field of a Reference 573 // object (either by using Unsafe directly or through reflection) 574 // then, if SATB is enabled, we need to record the referent in an 575 // SATB log buffer using the pre-barrier mechanism. 576 // Also we need to add memory barrier to prevent commoning reads 577 // from this field across safepoint since GC can change its value. 578 if (!on_weak_ref || (unknown && (offset == top || obj == top)) || !keep_alive) { 579 return load; 580 } 581 582 assert(access.is_parse_access(), "entry not supported at optimization time"); 583 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access); 584 GraphKit* kit = parse_access.kit(); 585 bool mismatched = (decorators & C2_MISMATCHED) != 0; 586 bool is_unordered = (decorators & MO_UNORDERED) != 0; 587 bool in_native = (decorators & IN_NATIVE) != 0; 588 bool need_cpu_mem_bar = !is_unordered || mismatched || in_native; 589 590 if (on_weak_ref) { 591 // Use the pre-barrier to record the value in the referent field 592 satb_write_barrier_pre(kit, false /* do_load */, 593 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 594 load /* pre_val */, T_OBJECT); 595 // Add memory barrier to prevent commoning reads from this field 596 // across safepoint since GC can change its value. 597 kit->insert_mem_bar(Op_MemBarCPUOrder); 598 } else if (unknown) { 599 // We do not require a mem bar inside pre_barrier if need_mem_bar 600 // is set: the barriers would be emitted by us. 601 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar); 602 } 603 } 604 605 return load; 606 } 607 608 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val, 609 Node* new_val, const Type* value_type) const { 610 GraphKit* kit = access.kit(); 611 if (access.is_oop()) { 612 new_val = shenandoah_storeval_barrier(kit, new_val); 613 shenandoah_write_barrier_pre(kit, false /* do_load */, 614 NULL, NULL, max_juint, NULL, NULL, 615 expected_val /* pre_val */, T_OBJECT); 616 617 MemNode::MemOrd mo = access.mem_node_mo(); 618 Node* mem = access.memory(); 619 Node* adr = access.addr().node(); 620 const TypePtr* adr_type = access.addr().type(); 621 Node* load_store = NULL; 622 623 #ifdef _LP64 624 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 625 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop())); 626 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop())); 627 if (ShenandoahCASBarrier) { 628 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo)); 629 } else { 630 load_store = kit->gvn().transform(new CompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo)); 631 } 632 } else 633 #endif 634 { 635 if (ShenandoahCASBarrier) { 636 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo)); 637 } else { 638 load_store = kit->gvn().transform(new CompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo)); 639 } 640 } 641 642 access.set_raw_access(load_store); 643 pin_atomic_op(access); 644 645 #ifdef _LP64 646 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 647 load_store = kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 648 } 649 #endif 650 load_store = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, load_store, false)); 651 return load_store; 652 } 653 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type); 654 } 655 656 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val, 657 Node* new_val, const Type* value_type) const { 658 GraphKit* kit = access.kit(); 659 if (access.is_oop()) { 660 new_val = shenandoah_storeval_barrier(kit, new_val); 661 shenandoah_write_barrier_pre(kit, false /* do_load */, 662 NULL, NULL, max_juint, NULL, NULL, 663 expected_val /* pre_val */, T_OBJECT); 664 DecoratorSet decorators = access.decorators(); 665 MemNode::MemOrd mo = access.mem_node_mo(); 666 Node* mem = access.memory(); 667 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0; 668 Node* load_store = NULL; 669 Node* adr = access.addr().node(); 670 #ifdef _LP64 671 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 672 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop())); 673 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop())); 674 if (ShenandoahCASBarrier) { 675 if (is_weak_cas) { 676 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 677 } else { 678 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 679 } 680 } else { 681 if (is_weak_cas) { 682 load_store = kit->gvn().transform(new WeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 683 } else { 684 load_store = kit->gvn().transform(new CompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 685 } 686 } 687 } else 688 #endif 689 { 690 if (ShenandoahCASBarrier) { 691 if (is_weak_cas) { 692 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 693 } else { 694 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 695 } 696 } else { 697 if (is_weak_cas) { 698 load_store = kit->gvn().transform(new WeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 699 } else { 700 load_store = kit->gvn().transform(new CompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 701 } 702 } 703 } 704 access.set_raw_access(load_store); 705 pin_atomic_op(access); 706 return load_store; 707 } 708 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type); 709 } 710 711 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const { 712 GraphKit* kit = access.kit(); 713 if (access.is_oop()) { 714 val = shenandoah_storeval_barrier(kit, val); 715 } 716 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type); 717 if (access.is_oop()) { 718 result = kit->gvn().transform(new ShenandoahLoadReferenceBarrierNode(NULL, result, false)); 719 shenandoah_write_barrier_pre(kit, false /* do_load */, 720 NULL, NULL, max_juint, NULL, NULL, 721 result /* pre_val */, T_OBJECT); 722 } 723 return result; 724 } 725 726 // Support for GC barriers emitted during parsing 727 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const { 728 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) return true; 729 if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) { 730 return false; 731 } 732 CallLeafNode *call = node->as_CallLeaf(); 733 if (call->_name == NULL) { 734 return false; 735 } 736 737 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 || 738 strcmp(call->_name, "shenandoah_cas_obj") == 0 || 739 strcmp(call->_name, "shenandoah_wb_pre") == 0; 740 } 741 742 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const { 743 if (c == NULL) { 744 return c; 745 } 746 if (c->Opcode() == Op_ShenandoahLoadReferenceBarrier) { 747 return c->in(ShenandoahLoadReferenceBarrierNode::ValueIn); 748 } 749 if (c->Opcode() == Op_ShenandoahEnqueueBarrier) { 750 c = c->in(1); 751 } 752 return c; 753 } 754 755 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const { 756 return !ShenandoahBarrierC2Support::expand(C, igvn); 757 } 758 759 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const { 760 if (mode == LoopOptsShenandoahExpand) { 761 assert(UseShenandoahGC, "only for shenandoah"); 762 ShenandoahBarrierC2Support::pin_and_expand(phase); 763 return true; 764 } else if (mode == LoopOptsShenandoahPostExpand) { 765 assert(UseShenandoahGC, "only for shenandoah"); 766 visited.clear(); 767 ShenandoahBarrierC2Support::optimize_after_expansion(visited, nstack, worklist, phase); 768 return true; 769 } 770 return false; 771 } 772 773 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, ArrayCopyPhase phase) const { 774 bool is_oop = is_reference_type(type); 775 if (!is_oop) { 776 return false; 777 } 778 if (tightly_coupled_alloc) { 779 if (phase == Optimization) { 780 return false; 781 } 782 return !is_clone; 783 } 784 if (phase == Optimization) { 785 return !ShenandoahStoreValEnqueueBarrier; 786 } 787 return true; 788 } 789 790 bool ShenandoahBarrierSetC2::clone_needs_barrier(Node* src, PhaseGVN& gvn) { 791 const TypeOopPtr* src_type = gvn.type(src)->is_oopptr(); 792 if (src_type->isa_instptr() != NULL) { 793 ciInstanceKlass* ik = src_type->klass()->as_instance_klass(); 794 if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) { 795 if (ik->has_object_fields()) { 796 return true; 797 } else { 798 if (!src_type->klass_is_exact()) { 799 Compile::current()->dependencies()->assert_leaf_type(ik); 800 } 801 } 802 } else { 803 return true; 804 } 805 } else if (src_type->isa_aryptr()) { 806 BasicType src_elem = src_type->klass()->as_array_klass()->element_type()->basic_type(); 807 if (is_reference_type(src_elem)) { 808 return true; 809 } 810 } else { 811 return true; 812 } 813 return false; 814 } 815 816 void ShenandoahBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const { 817 Node* ctrl = ac->in(TypeFunc::Control); 818 Node* mem = ac->in(TypeFunc::Memory); 819 Node* src_base = ac->in(ArrayCopyNode::Src); 820 Node* src_offset = ac->in(ArrayCopyNode::SrcPos); 821 Node* dest_base = ac->in(ArrayCopyNode::Dest); 822 Node* dest_offset = ac->in(ArrayCopyNode::DestPos); 823 Node* length = ac->in(ArrayCopyNode::Length); 824 825 Node* src = phase->basic_plus_adr(src_base, src_offset); 826 Node* dest = phase->basic_plus_adr(dest_base, dest_offset); 827 828 if (ShenandoahCloneBarrier && clone_needs_barrier(src, phase->igvn())) { 829 // Check if heap is has forwarded objects. If it does, we need to call into the special 830 // routine that would fix up source references before we can continue. 831 832 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT }; 833 Node* region = new RegionNode(PATH_LIMIT); 834 Node* mem_phi = new PhiNode(region, Type::MEMORY, TypeRawPtr::BOTTOM); 835 836 Node* thread = phase->transform_later(new ThreadLocalNode()); 837 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())); 838 Node* gc_state_addr = phase->transform_later(new AddPNode(phase->C->top(), thread, offset)); 839 840 uint gc_state_idx = Compile::AliasIdxRaw; 841 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument 842 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx)); 843 844 Node* gc_state = phase->transform_later(new LoadBNode(ctrl, mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered)); 845 Node* stable_and = phase->transform_later(new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED))); 846 Node* stable_cmp = phase->transform_later(new CmpINode(stable_and, phase->igvn().zerocon(T_INT))); 847 Node* stable_test = phase->transform_later(new BoolNode(stable_cmp, BoolTest::ne)); 848 849 IfNode* stable_iff = phase->transform_later(new IfNode(ctrl, stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN))->as_If(); 850 Node* stable_ctrl = phase->transform_later(new IfFalseNode(stable_iff)); 851 Node* unstable_ctrl = phase->transform_later(new IfTrueNode(stable_iff)); 852 853 // Heap is stable, no need to do anything additional 854 region->init_req(_heap_stable, stable_ctrl); 855 mem_phi->init_req(_heap_stable, mem); 856 857 // Heap is unstable, call into clone barrier stub 858 Node* call = phase->make_leaf_call(unstable_ctrl, mem, 859 ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(), 860 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier), 861 "shenandoah_clone", 862 TypeRawPtr::BOTTOM, 863 src_base); 864 call = phase->transform_later(call); 865 866 ctrl = phase->transform_later(new ProjNode(call, TypeFunc::Control)); 867 mem = phase->transform_later(new ProjNode(call, TypeFunc::Memory)); 868 region->init_req(_heap_unstable, ctrl); 869 mem_phi->init_req(_heap_unstable, mem); 870 871 // Wire up the actual arraycopy stub now 872 ctrl = phase->transform_later(region); 873 mem = phase->transform_later(mem_phi); 874 875 const char* name = "arraycopy"; 876 call = phase->make_leaf_call(ctrl, mem, 877 OptoRuntime::fast_arraycopy_Type(), 878 phase->basictype2arraycopy(T_LONG, NULL, NULL, true, name, true), 879 name, TypeRawPtr::BOTTOM, 880 src, dest, length 881 LP64_ONLY(COMMA phase->top())); 882 call = phase->transform_later(call); 883 884 // Hook up the whole thing into the graph 885 phase->igvn().replace_node(ac, call); 886 } else { 887 BarrierSetC2::clone_at_expansion(phase, ac); 888 } 889 } 890 891 892 // Support for macro expanded GC barriers 893 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const { 894 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) { 895 state()->add_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node); 896 } 897 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) { 898 state()->add_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node); 899 } 900 } 901 902 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const { 903 if (node->Opcode() == Op_ShenandoahEnqueueBarrier) { 904 state()->remove_enqueue_barrier((ShenandoahEnqueueBarrierNode*) node); 905 } 906 if (node->Opcode() == Op_ShenandoahLoadReferenceBarrier) { 907 state()->remove_load_reference_barrier((ShenandoahLoadReferenceBarrierNode*) node); 908 } 909 } 910 911 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const { 912 if (is_shenandoah_wb_pre_call(n)) { 913 shenandoah_eliminate_wb_pre(n, ¯o->igvn()); 914 } 915 } 916 917 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const { 918 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), ""); 919 Node* c = call->as_Call()->proj_out(TypeFunc::Control); 920 c = c->unique_ctrl_out(); 921 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 922 c = c->unique_ctrl_out(); 923 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 924 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0); 925 assert(iff->is_If(), "expect test"); 926 if (!is_shenandoah_marking_if(igvn, iff)) { 927 c = c->unique_ctrl_out(); 928 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 929 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0); 930 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test"); 931 } 932 Node* cmpx = iff->in(1)->in(1); 933 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ)); 934 igvn->rehash_node_delayed(call); 935 call->del_req(call->req()-1); 936 } 937 938 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const { 939 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) { 940 igvn->add_users_to_worklist(node); 941 } 942 } 943 944 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const { 945 for (uint i = 0; i < useful.size(); i++) { 946 Node* n = useful.at(i); 947 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) { 948 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 949 C->record_for_igvn(n->fast_out(i)); 950 } 951 } 952 } 953 for (int i = state()->enqueue_barriers_count() - 1; i >= 0; i--) { 954 ShenandoahEnqueueBarrierNode* n = state()->enqueue_barrier(i); 955 if (!useful.member(n)) { 956 state()->remove_enqueue_barrier(n); 957 } 958 } 959 for (int i = state()->load_reference_barriers_count() - 1; i >= 0; i--) { 960 ShenandoahLoadReferenceBarrierNode* n = state()->load_reference_barrier(i); 961 if (!useful.member(n)) { 962 state()->remove_load_reference_barrier(n); 963 } 964 } 965 } 966 967 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const { 968 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena); 969 } 970 971 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const { 972 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state()); 973 } 974 975 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be 976 // expanded later, then now is the time to do so. 977 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; } 978 979 #ifdef ASSERT 980 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const { 981 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeMacroExpand) { 982 ShenandoahBarrierC2Support::verify(Compile::current()->root()); 983 } else if (phase == BarrierSetC2::BeforeCodeGen) { 984 // Verify G1 pre-barriers 985 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()); 986 987 ResourceArea *area = Thread::current()->resource_area(); 988 Unique_Node_List visited(area); 989 Node_List worklist(area); 990 // We're going to walk control flow backwards starting from the Root 991 worklist.push(compile->root()); 992 while (worklist.size() > 0) { 993 Node *x = worklist.pop(); 994 if (x == NULL || x == compile->top()) continue; 995 if (visited.member(x)) { 996 continue; 997 } else { 998 visited.push(x); 999 } 1000 1001 if (x->is_Region()) { 1002 for (uint i = 1; i < x->req(); i++) { 1003 worklist.push(x->in(i)); 1004 } 1005 } else { 1006 worklist.push(x->in(0)); 1007 // We are looking for the pattern: 1008 // /->ThreadLocal 1009 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 1010 // \->ConI(0) 1011 // We want to verify that the If and the LoadB have the same control 1012 // See GraphKit::g1_write_barrier_pre() 1013 if (x->is_If()) { 1014 IfNode *iff = x->as_If(); 1015 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 1016 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 1017 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 1018 && cmp->in(1)->is_Load()) { 1019 LoadNode *load = cmp->in(1)->as_Load(); 1020 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 1021 && load->in(2)->in(3)->is_Con() 1022 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 1023 1024 Node *if_ctrl = iff->in(0); 1025 Node *load_ctrl = load->in(0); 1026 1027 if (if_ctrl != load_ctrl) { 1028 // Skip possible CProj->NeverBranch in infinite loops 1029 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 1030 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 1031 if_ctrl = if_ctrl->in(0)->in(0); 1032 } 1033 } 1034 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 1035 } 1036 } 1037 } 1038 } 1039 } 1040 } 1041 } 1042 } 1043 #endif 1044 1045 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const { 1046 if (is_shenandoah_wb_pre_call(n)) { 1047 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt(); 1048 if (n->req() > cnt) { 1049 Node* addp = n->in(cnt); 1050 if (has_only_shenandoah_wb_pre_uses(addp)) { 1051 n->del_req(cnt); 1052 if (can_reshape) { 1053 phase->is_IterGVN()->_worklist.push(addp); 1054 } 1055 return n; 1056 } 1057 } 1058 } 1059 if (n->Opcode() == Op_CmpP) { 1060 Node* in1 = n->in(1); 1061 Node* in2 = n->in(2); 1062 if (in1->bottom_type() == TypePtr::NULL_PTR) { 1063 in2 = step_over_gc_barrier(in2); 1064 } 1065 if (in2->bottom_type() == TypePtr::NULL_PTR) { 1066 in1 = step_over_gc_barrier(in1); 1067 } 1068 PhaseIterGVN* igvn = phase->is_IterGVN(); 1069 if (in1 != n->in(1)) { 1070 if (igvn != NULL) { 1071 n->set_req_X(1, in1, igvn); 1072 } else { 1073 n->set_req(1, in1); 1074 } 1075 assert(in2 == n->in(2), "only one change"); 1076 return n; 1077 } 1078 if (in2 != n->in(2)) { 1079 if (igvn != NULL) { 1080 n->set_req_X(2, in2, igvn); 1081 } else { 1082 n->set_req(2, in2); 1083 } 1084 return n; 1085 } 1086 } else if (can_reshape && 1087 n->Opcode() == Op_If && 1088 ShenandoahBarrierC2Support::is_heap_stable_test(n) && 1089 n->in(0) != NULL) { 1090 Node* dom = n->in(0); 1091 Node* prev_dom = n; 1092 int op = n->Opcode(); 1093 int dist = 16; 1094 // Search up the dominator tree for another heap stable test 1095 while (dom->Opcode() != op || // Not same opcode? 1096 !ShenandoahBarrierC2Support::is_heap_stable_test(dom) || // Not same input 1? 1097 prev_dom->in(0) != dom) { // One path of test does not dominate? 1098 if (dist < 0) return NULL; 1099 1100 dist--; 1101 prev_dom = dom; 1102 dom = IfNode::up_one_dom(dom); 1103 if (!dom) return NULL; 1104 } 1105 1106 // Check that we did not follow a loop back to ourselves 1107 if (n == dom) { 1108 return NULL; 1109 } 1110 1111 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN()); 1112 } 1113 1114 return NULL; 1115 } 1116 1117 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) { 1118 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1119 Node* u = n->fast_out(i); 1120 if (!is_shenandoah_wb_pre_call(u)) { 1121 return false; 1122 } 1123 } 1124 return n->outcnt() > 0; 1125 } 1126 1127 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode) const { 1128 switch (opcode) { 1129 case Op_CallLeaf: 1130 case Op_CallLeafNoFP: { 1131 assert (n->is_Call(), ""); 1132 CallNode *call = n->as_Call(); 1133 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) { 1134 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt(); 1135 if (call->req() > cnt) { 1136 assert(call->req() == cnt + 1, "only one extra input"); 1137 Node *addp = call->in(cnt); 1138 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?"); 1139 call->del_req(cnt); 1140 } 1141 } 1142 return false; 1143 } 1144 case Op_ShenandoahCompareAndSwapP: 1145 case Op_ShenandoahCompareAndSwapN: 1146 case Op_ShenandoahWeakCompareAndSwapN: 1147 case Op_ShenandoahWeakCompareAndSwapP: 1148 case Op_ShenandoahCompareAndExchangeP: 1149 case Op_ShenandoahCompareAndExchangeN: 1150 #ifdef ASSERT 1151 if( VerifyOptoOopOffsets ) { 1152 MemNode* mem = n->as_Mem(); 1153 // Check to see if address types have grounded out somehow. 1154 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 1155 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 1156 bool oop_offset_is_sane = k->contains_field_offset(tp->offset()); 1157 assert( !tp || oop_offset_is_sane, "" ); 1158 } 1159 #endif 1160 return true; 1161 case Op_ShenandoahLoadReferenceBarrier: 1162 assert(false, "should have been expanded already"); 1163 return true; 1164 default: 1165 return false; 1166 } 1167 } 1168 1169 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const { 1170 switch (opcode) { 1171 case Op_ShenandoahCompareAndExchangeP: 1172 case Op_ShenandoahCompareAndExchangeN: 1173 conn_graph->add_objload_to_connection_graph(n, delayed_worklist); 1174 // fallthrough 1175 case Op_ShenandoahWeakCompareAndSwapP: 1176 case Op_ShenandoahWeakCompareAndSwapN: 1177 case Op_ShenandoahCompareAndSwapP: 1178 case Op_ShenandoahCompareAndSwapN: 1179 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist); 1180 return true; 1181 case Op_StoreP: { 1182 Node* adr = n->in(MemNode::Address); 1183 const Type* adr_type = gvn->type(adr); 1184 // Pointer stores in G1 barriers looks like unsafe access. 1185 // Ignore such stores to be able scalar replace non-escaping 1186 // allocations. 1187 if (adr_type->isa_rawptr() && adr->is_AddP()) { 1188 Node* base = conn_graph->get_addp_base(adr); 1189 if (base->Opcode() == Op_LoadP && 1190 base->in(MemNode::Address)->is_AddP()) { 1191 adr = base->in(MemNode::Address); 1192 Node* tls = conn_graph->get_addp_base(adr); 1193 if (tls->Opcode() == Op_ThreadLocal) { 1194 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 1195 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()); 1196 if (offs == buf_offset) { 1197 return true; // Pre barrier previous oop value store. 1198 } 1199 } 1200 } 1201 } 1202 return false; 1203 } 1204 case Op_ShenandoahEnqueueBarrier: 1205 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist); 1206 break; 1207 case Op_ShenandoahLoadReferenceBarrier: 1208 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), delayed_worklist); 1209 return true; 1210 default: 1211 // Nothing 1212 break; 1213 } 1214 return false; 1215 } 1216 1217 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const { 1218 switch (opcode) { 1219 case Op_ShenandoahCompareAndExchangeP: 1220 case Op_ShenandoahCompareAndExchangeN: { 1221 Node *adr = n->in(MemNode::Address); 1222 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 1223 // fallthrough 1224 } 1225 case Op_ShenandoahCompareAndSwapP: 1226 case Op_ShenandoahCompareAndSwapN: 1227 case Op_ShenandoahWeakCompareAndSwapP: 1228 case Op_ShenandoahWeakCompareAndSwapN: 1229 return conn_graph->add_final_edges_unsafe_access(n, opcode); 1230 case Op_ShenandoahEnqueueBarrier: 1231 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), NULL); 1232 return true; 1233 case Op_ShenandoahLoadReferenceBarrier: 1234 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahLoadReferenceBarrierNode::ValueIn), NULL); 1235 return true; 1236 default: 1237 // Nothing 1238 break; 1239 } 1240 return false; 1241 } 1242 1243 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const { 1244 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) || 1245 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN); 1246 1247 } 1248 1249 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const { 1250 switch (opcode) { 1251 case Op_ShenandoahCompareAndExchangeP: 1252 case Op_ShenandoahCompareAndExchangeN: 1253 case Op_ShenandoahWeakCompareAndSwapP: 1254 case Op_ShenandoahWeakCompareAndSwapN: 1255 case Op_ShenandoahCompareAndSwapP: 1256 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree 1257 Node* newval = n->in(MemNode::ValueIn); 1258 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn); 1259 Node* pair = new BinaryNode(oldval, newval); 1260 n->set_req(MemNode::ValueIn,pair); 1261 n->del_req(LoadStoreConditionalNode::ExpectedIn); 1262 return true; 1263 } 1264 default: 1265 break; 1266 } 1267 return false; 1268 } 1269 1270 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const { 1271 return xop == Op_ShenandoahCompareAndExchangeP || 1272 xop == Op_ShenandoahCompareAndExchangeN || 1273 xop == Op_ShenandoahWeakCompareAndSwapP || 1274 xop == Op_ShenandoahWeakCompareAndSwapN || 1275 xop == Op_ShenandoahCompareAndSwapN || 1276 xop == Op_ShenandoahCompareAndSwapP; 1277 }