1 /* 2 * Copyright (c) 2018, Red Hat, Inc. and/or its affiliates. 3 * 4 * This code is free software; you can redistribute it and/or modify it 5 * under the terms of the GNU General Public License version 2 only, as 6 * published by the Free Software Foundation. 7 * 8 * This code is distributed in the hope that it will be useful, but WITHOUT 9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 11 * version 2 for more details (a copy is included in the LICENSE file that 12 * accompanied this code). 13 * 14 * You should have received a copy of the GNU General Public License version 15 * 2 along with this work; if not, write to the Free Software Foundation, 16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 17 * 18 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 19 * or visit www.oracle.com if you need additional information or have any 20 * questions. 21 * 22 */ 23 24 #include "precompiled.hpp" 25 #include "gc/shenandoah/shenandoahHeap.hpp" 26 #include "gc/shenandoah/shenandoahHeuristics.hpp" 27 #include "gc/shenandoah/shenandoahRuntime.hpp" 28 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp" 29 #include "gc/shenandoah/c2/shenandoahSupport.hpp" 30 #include "opto/arraycopynode.hpp" 31 #include "opto/graphKit.hpp" 32 #include "opto/idealKit.hpp" 33 #include "opto/macro.hpp" 34 #include "opto/narrowptrnode.hpp" 35 36 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() { 37 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2()); 38 } 39 40 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena) 41 : _shenandoah_barriers(new (comp_arena) GrowableArray<ShenandoahWriteBarrierNode*>(comp_arena, 8, 0, NULL)) { 42 } 43 44 int ShenandoahBarrierSetC2State::shenandoah_barriers_count() const { 45 return _shenandoah_barriers->length(); 46 } 47 48 ShenandoahWriteBarrierNode* ShenandoahBarrierSetC2State::shenandoah_barrier(int idx) const { 49 return _shenandoah_barriers->at(idx); 50 } 51 52 void ShenandoahBarrierSetC2State::add_shenandoah_barrier(ShenandoahWriteBarrierNode * n) { 53 assert(!_shenandoah_barriers->contains(n), "duplicate entry in barrier list"); 54 _shenandoah_barriers->append(n); 55 } 56 57 void ShenandoahBarrierSetC2State::remove_shenandoah_barrier(ShenandoahWriteBarrierNode * n) { 58 if (_shenandoah_barriers->contains(n)) { 59 _shenandoah_barriers->remove(n); 60 } 61 } 62 63 #define __ kit-> 64 65 Node* ShenandoahBarrierSetC2::shenandoah_read_barrier(GraphKit* kit, Node* obj) const { 66 if (ShenandoahReadBarrier) { 67 obj = shenandoah_read_barrier_impl(kit, obj, false, true, true); 68 } 69 return obj; 70 } 71 72 Node* ShenandoahBarrierSetC2::shenandoah_storeval_barrier(GraphKit* kit, Node* obj) const { 73 if (ShenandoahStoreValEnqueueBarrier) { 74 obj = shenandoah_write_barrier(kit, obj); 75 obj = shenandoah_enqueue_barrier(kit, obj); 76 } 77 if (ShenandoahStoreValReadBarrier) { 78 obj = shenandoah_read_barrier_impl(kit, obj, true, false, false); 79 } 80 return obj; 81 } 82 83 Node* ShenandoahBarrierSetC2::shenandoah_read_barrier_acmp(GraphKit* kit, Node* obj) { 84 return shenandoah_read_barrier_impl(kit, obj, true, true, false); 85 } 86 87 Node* ShenandoahBarrierSetC2::shenandoah_read_barrier_impl(GraphKit* kit, Node* obj, bool use_ctrl, bool use_mem, bool allow_fromspace) const { 88 const Type* obj_type = obj->bottom_type(); 89 if (obj_type->higher_equal(TypePtr::NULL_PTR)) { 90 return obj; 91 } 92 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type); 93 Node* mem = use_mem ? __ memory(adr_type) : __ immutable_memory(); 94 95 if (! ShenandoahBarrierNode::needs_barrier(&__ gvn(), NULL, obj, mem, allow_fromspace)) { 96 // We know it is null, no barrier needed. 97 return obj; 98 } 99 100 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) { 101 102 // We don't know if it's null or not. Need null-check. 103 enum { _not_null_path = 1, _null_path, PATH_LIMIT }; 104 RegionNode* region = new RegionNode(PATH_LIMIT); 105 Node* phi = new PhiNode(region, obj_type); 106 Node* null_ctrl = __ top(); 107 Node* not_null_obj = __ null_check_oop(obj, &null_ctrl); 108 109 region->init_req(_null_path, null_ctrl); 110 phi ->init_req(_null_path, __ zerocon(T_OBJECT)); 111 112 Node* ctrl = use_ctrl ? __ control() : NULL; 113 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, not_null_obj, allow_fromspace); 114 Node* n = __ gvn().transform(rb); 115 116 region->init_req(_not_null_path, __ control()); 117 phi ->init_req(_not_null_path, n); 118 119 __ set_control(__ gvn().transform(region)); 120 __ record_for_igvn(region); 121 return __ gvn().transform(phi); 122 123 } else { 124 // We know it is not null. Simple barrier is sufficient. 125 Node* ctrl = use_ctrl ? __ control() : NULL; 126 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, obj, allow_fromspace); 127 Node* n = __ gvn().transform(rb); 128 __ record_for_igvn(n); 129 return n; 130 } 131 } 132 133 Node* ShenandoahBarrierSetC2::shenandoah_write_barrier_helper(GraphKit* kit, Node* obj, const TypePtr* adr_type) const { 134 ShenandoahWriteBarrierNode* wb = new ShenandoahWriteBarrierNode(kit->C, kit->control(), kit->memory(adr_type), obj); 135 Node* n = __ gvn().transform(wb); 136 if (n == wb) { // New barrier needs memory projection. 137 Node* proj = __ gvn().transform(new ShenandoahWBMemProjNode(n)); 138 __ set_memory(proj, adr_type); 139 } 140 return n; 141 } 142 143 Node* ShenandoahBarrierSetC2::shenandoah_write_barrier(GraphKit* kit, Node* obj) const { 144 if (ShenandoahWriteBarrier) { 145 obj = shenandoah_write_barrier_impl(kit, obj); 146 } 147 return obj; 148 } 149 150 Node* ShenandoahBarrierSetC2::shenandoah_write_barrier_impl(GraphKit* kit, Node* obj) const { 151 if (! ShenandoahBarrierNode::needs_barrier(&__ gvn(), NULL, obj, NULL, true)) { 152 return obj; 153 } 154 const Type* obj_type = obj->bottom_type(); 155 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type); 156 Node* n = shenandoah_write_barrier_helper(kit, obj, adr_type); 157 __ record_for_igvn(n); 158 return n; 159 } 160 161 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseTransform* phase, Node* adr, 162 BasicType bt, uint adr_idx) const { 163 intptr_t offset = 0; 164 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 165 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 166 167 if (offset == Type::OffsetBot) { 168 return false; // cannot unalias unless there are precise offsets 169 } 170 171 if (alloc == NULL) { 172 return false; // No allocation found 173 } 174 175 intptr_t size_in_bytes = type2aelembytes(bt); 176 177 Node* mem = __ memory(adr_idx); // start searching here... 178 179 for (int cnt = 0; cnt < 50; cnt++) { 180 181 if (mem->is_Store()) { 182 183 Node* st_adr = mem->in(MemNode::Address); 184 intptr_t st_offset = 0; 185 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 186 187 if (st_base == NULL) { 188 break; // inscrutable pointer 189 } 190 191 // Break we have found a store with same base and offset as ours so break 192 if (st_base == base && st_offset == offset) { 193 break; 194 } 195 196 if (st_offset != offset && st_offset != Type::OffsetBot) { 197 const int MAX_STORE = BytesPerLong; 198 if (st_offset >= offset + size_in_bytes || 199 st_offset <= offset - MAX_STORE || 200 st_offset <= offset - mem->as_Store()->memory_size()) { 201 // Success: The offsets are provably independent. 202 // (You may ask, why not just test st_offset != offset and be done? 203 // The answer is that stores of different sizes can co-exist 204 // in the same sequence of RawMem effects. We sometimes initialize 205 // a whole 'tile' of array elements with a single jint or jlong.) 206 mem = mem->in(MemNode::Memory); 207 continue; // advance through independent store memory 208 } 209 } 210 211 if (st_base != base 212 && MemNode::detect_ptr_independence(base, alloc, st_base, 213 AllocateNode::Ideal_allocation(st_base, phase), 214 phase)) { 215 // Success: The bases are provably independent. 216 mem = mem->in(MemNode::Memory); 217 continue; // advance through independent store memory 218 } 219 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 220 221 InitializeNode* st_init = mem->in(0)->as_Initialize(); 222 AllocateNode* st_alloc = st_init->allocation(); 223 224 // Make sure that we are looking at the same allocation site. 225 // The alloc variable is guaranteed to not be null here from earlier check. 226 if (alloc == st_alloc) { 227 // Check that the initialization is storing NULL so that no previous store 228 // has been moved up and directly write a reference 229 Node* captured_store = st_init->find_captured_store(offset, 230 type2aelembytes(T_OBJECT), 231 phase); 232 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 233 return true; 234 } 235 } 236 } 237 238 // Unless there is an explicit 'continue', we must bail out here, 239 // because 'mem' is an inscrutable memory state (e.g., a call). 240 break; 241 } 242 243 return false; 244 } 245 246 #undef __ 247 #define __ ideal. 248 249 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit, 250 bool do_load, 251 Node* obj, 252 Node* adr, 253 uint alias_idx, 254 Node* val, 255 const TypeOopPtr* val_type, 256 Node* pre_val, 257 BasicType bt) const { 258 // Some sanity checks 259 // Note: val is unused in this routine. 260 261 if (do_load) { 262 // We need to generate the load of the previous value 263 assert(obj != NULL, "must have a base"); 264 assert(adr != NULL, "where are loading from?"); 265 assert(pre_val == NULL, "loaded already?"); 266 assert(val_type != NULL, "need a type"); 267 268 if (ReduceInitialCardMarks 269 && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) { 270 return; 271 } 272 273 } else { 274 // In this case both val_type and alias_idx are unused. 275 assert(pre_val != NULL, "must be loaded already"); 276 // Nothing to be done if pre_val is null. 277 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 278 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 279 } 280 assert(bt == T_OBJECT, "or we shouldn't be here"); 281 282 IdealKit ideal(kit, true); 283 284 Node* tls = __ thread(); // ThreadLocalStorage 285 286 Node* no_base = __ top(); 287 Node* zero = __ ConI(0); 288 Node* zeroX = __ ConX(0); 289 290 float likely = PROB_LIKELY(0.999); 291 float unlikely = PROB_UNLIKELY(0.999); 292 293 // Offsets into the thread 294 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()); 295 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()); 296 297 // Now the actual pointers into the thread 298 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 299 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 300 301 // Now some of the values 302 Node* marking; 303 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()))); 304 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw); 305 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING)); 306 assert(ShenandoahWriteBarrierNode::is_gc_state_load(ld), "Should match the shape"); 307 308 // if (!marking) 309 __ if_then(marking, BoolTest::ne, zero, unlikely); { 310 BasicType index_bt = TypeX_X->basic_type(); 311 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 312 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 313 314 if (do_load) { 315 // load original value 316 // alias_idx correct?? 317 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 318 } 319 320 // if (pre_val != NULL) 321 __ if_then(pre_val, BoolTest::ne, kit->null()); { 322 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 323 324 // is the queue for this thread full? 325 __ if_then(index, BoolTest::ne, zeroX, likely); { 326 327 // decrement the index 328 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 329 330 // Now get the buffer location we will log the previous value into and store it 331 Node *log_addr = __ AddP(no_base, buffer, next_index); 332 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 333 // update the index 334 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 335 336 } __ else_(); { 337 338 // logging buffer is full, call the runtime 339 const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type(); 340 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls); 341 } __ end_if(); // (!index) 342 } __ end_if(); // (pre_val != NULL) 343 } __ end_if(); // (!marking) 344 345 // Final sync IdealKit and GraphKit. 346 kit->final_sync(ideal); 347 348 if (ShenandoahSATBBarrier && adr != NULL) { 349 Node* c = kit->control(); 350 Node* call = c->in(1)->in(1)->in(1)->in(0); 351 assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected"); 352 call->add_req(adr); 353 } 354 } 355 356 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) { 357 return call->is_CallLeaf() && 358 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry); 359 } 360 361 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseTransform *phase, Node* n) { 362 if (n->Opcode() != Op_If) { 363 return false; 364 } 365 366 Node* bol = n->in(1); 367 assert(bol->is_Bool(), ""); 368 Node* cmpx = bol->in(1); 369 if (bol->as_Bool()->_test._test == BoolTest::ne && 370 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) && 371 is_shenandoah_state_load(cmpx->in(1)->in(1)) && 372 cmpx->in(1)->in(2)->is_Con() && 373 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) { 374 return true; 375 } 376 377 return false; 378 } 379 380 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) { 381 if (!n->is_Load()) return false; 382 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset()); 383 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal 384 && n->in(2)->in(3)->is_Con() 385 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset; 386 } 387 388 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit, 389 bool do_load, 390 Node* obj, 391 Node* adr, 392 uint alias_idx, 393 Node* val, 394 const TypeOopPtr* val_type, 395 Node* pre_val, 396 BasicType bt) const { 397 if (ShenandoahSATBBarrier) { 398 IdealKit ideal(kit); 399 kit->sync_kit(ideal); 400 401 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt); 402 403 ideal.sync_kit(kit); 404 kit->final_sync(ideal); 405 } 406 } 407 408 Node* ShenandoahBarrierSetC2::shenandoah_enqueue_barrier(GraphKit* kit, Node* pre_val) const { 409 return kit->gvn().transform(new ShenandoahEnqueueBarrierNode(pre_val)); 410 } 411 412 // Helper that guards and inserts a pre-barrier. 413 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset, 414 Node* pre_val, bool need_mem_bar) const { 415 // We could be accessing the referent field of a reference object. If so, when G1 416 // is enabled, we need to log the value in the referent field in an SATB buffer. 417 // This routine performs some compile time filters and generates suitable 418 // runtime filters that guard the pre-barrier code. 419 // Also add memory barrier for non volatile load from the referent field 420 // to prevent commoning of loads across safepoint. 421 422 // Some compile time checks. 423 424 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 425 const TypeX* otype = offset->find_intptr_t_type(); 426 if (otype != NULL && otype->is_con() && 427 otype->get_con() != java_lang_ref_Reference::referent_offset) { 428 // Constant offset but not the reference_offset so just return 429 return; 430 } 431 432 // We only need to generate the runtime guards for instances. 433 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 434 if (btype != NULL) { 435 if (btype->isa_aryptr()) { 436 // Array type so nothing to do 437 return; 438 } 439 440 const TypeInstPtr* itype = btype->isa_instptr(); 441 if (itype != NULL) { 442 // Can the klass of base_oop be statically determined to be 443 // _not_ a sub-class of Reference and _not_ Object? 444 ciKlass* klass = itype->klass(); 445 if ( klass->is_loaded() && 446 !klass->is_subtype_of(kit->env()->Reference_klass()) && 447 !kit->env()->Object_klass()->is_subtype_of(klass)) { 448 return; 449 } 450 } 451 } 452 453 // The compile time filters did not reject base_oop/offset so 454 // we need to generate the following runtime filters 455 // 456 // if (offset == java_lang_ref_Reference::_reference_offset) { 457 // if (instance_of(base, java.lang.ref.Reference)) { 458 // pre_barrier(_, pre_val, ...); 459 // } 460 // } 461 462 float likely = PROB_LIKELY( 0.999); 463 float unlikely = PROB_UNLIKELY(0.999); 464 465 IdealKit ideal(kit); 466 467 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 468 469 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 470 // Update graphKit memory and control from IdealKit. 471 kit->sync_kit(ideal); 472 473 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass())); 474 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con); 475 476 // Update IdealKit memory and control from graphKit. 477 __ sync_kit(kit); 478 479 Node* one = __ ConI(1); 480 // is_instof == 0 if base_oop == NULL 481 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 482 483 // Update graphKit from IdeakKit. 484 kit->sync_kit(ideal); 485 486 // Use the pre-barrier to record the value in the referent field 487 satb_write_barrier_pre(kit, false /* do_load */, 488 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 489 pre_val /* pre_val */, 490 T_OBJECT); 491 if (need_mem_bar) { 492 // Add memory barrier to prevent commoning reads from this field 493 // across safepoint since GC can change its value. 494 kit->insert_mem_bar(Op_MemBarCPUOrder); 495 } 496 // Update IdealKit from graphKit. 497 __ sync_kit(kit); 498 499 } __ end_if(); // _ref_type != ref_none 500 } __ end_if(); // offset == referent_offset 501 502 // Final sync IdealKit and GraphKit. 503 kit->final_sync(ideal); 504 } 505 506 #undef __ 507 508 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() { 509 const Type **fields = TypeTuple::fields(2); 510 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value 511 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread 512 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields); 513 514 // create result type (range) 515 fields = TypeTuple::fields(0); 516 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields); 517 518 return TypeFunc::make(domain, range); 519 } 520 521 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() { 522 const Type **fields = TypeTuple::fields(1); 523 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value 524 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields); 525 526 // create result type (range) 527 fields = TypeTuple::fields(0); 528 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields); 529 530 return TypeFunc::make(domain, range); 531 } 532 533 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_write_barrier_Type() { 534 const Type **fields = TypeTuple::fields(1); 535 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value 536 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields); 537 538 // create result type (range) 539 fields = TypeTuple::fields(1); 540 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; 541 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields); 542 543 return TypeFunc::make(domain, range); 544 } 545 546 void ShenandoahBarrierSetC2::resolve_address(C2Access& access) const { 547 const TypePtr* adr_type = access.addr().type(); 548 549 if ((access.decorators() & IN_NATIVE) == 0 && (adr_type->isa_instptr() || adr_type->isa_aryptr())) { 550 int off = adr_type->is_ptr()->offset(); 551 int base_off = adr_type->isa_instptr() ? instanceOopDesc::base_offset_in_bytes() : 552 arrayOopDesc::base_offset_in_bytes(adr_type->is_aryptr()->elem()->array_element_basic_type()); 553 assert(off != Type::OffsetTop, "unexpected offset"); 554 if (off == Type::OffsetBot || off >= base_off) { 555 DecoratorSet decorators = access.decorators(); 556 bool is_write = (decorators & C2_WRITE_ACCESS) != 0; 557 GraphKit* kit = access.kit(); 558 Node* adr = access.addr().node(); 559 assert(adr->is_AddP(), "unexpected address shape"); 560 Node* base = adr->in(AddPNode::Base); 561 562 if (is_write) { 563 base = shenandoah_write_barrier(kit, base); 564 } else { 565 if (adr_type->isa_instptr()) { 566 Compile* C = kit->C; 567 ciField* field = C->alias_type(adr_type)->field(); 568 569 // Insert read barrier for Shenandoah. 570 if (field != NULL && 571 ((ShenandoahOptimizeStaticFinals && field->is_static() && field->is_final()) || 572 (ShenandoahOptimizeInstanceFinals && !field->is_static() && field->is_final()) || 573 (ShenandoahOptimizeStableFinals && field->is_stable()))) { 574 // Skip the barrier for special fields 575 } else { 576 base = shenandoah_read_barrier(kit, base); 577 } 578 } else { 579 base = shenandoah_read_barrier(kit, base); 580 } 581 } 582 if (base != adr->in(AddPNode::Base)) { 583 Node* address = adr->in(AddPNode::Address); 584 585 if (address->is_AddP()) { 586 assert(address->in(AddPNode::Base) == adr->in(AddPNode::Base), "unexpected address shape"); 587 assert(!address->in(AddPNode::Address)->is_AddP(), "unexpected address shape"); 588 assert(address->in(AddPNode::Address) == adr->in(AddPNode::Base), "unexpected address shape"); 589 address = address->clone(); 590 address->set_req(AddPNode::Base, base); 591 address->set_req(AddPNode::Address, base); 592 address = kit->gvn().transform(address); 593 } else { 594 assert(address == adr->in(AddPNode::Base), "unexpected address shape"); 595 address = base; 596 } 597 adr = adr->clone(); 598 adr->set_req(AddPNode::Base, base); 599 adr->set_req(AddPNode::Address, address); 600 adr = kit->gvn().transform(adr); 601 access.addr().set_node(adr); 602 } 603 } 604 } 605 } 606 607 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const { 608 DecoratorSet decorators = access.decorators(); 609 GraphKit* kit = access.kit(); 610 611 const TypePtr* adr_type = access.addr().type(); 612 Node* adr = access.addr().node(); 613 614 bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0; 615 bool on_heap = (decorators & IN_HEAP) != 0; 616 617 if (!access.is_oop() || (!on_heap && !anonymous)) { 618 return BarrierSetC2::store_at_resolved(access, val); 619 } 620 621 uint adr_idx = kit->C->get_alias_index(adr_type); 622 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 623 Node* value = val.node(); 624 value = shenandoah_storeval_barrier(kit, value); 625 val.set_node(value); 626 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(), 627 static_cast<const TypeOopPtr*>(val.type()), NULL /* pre_val */, access.type()); 628 return BarrierSetC2::store_at_resolved(access, val); 629 } 630 631 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const { 632 DecoratorSet decorators = access.decorators(); 633 GraphKit* kit = access.kit(); 634 635 Node* adr = access.addr().node(); 636 Node* obj = access.base(); 637 638 bool mismatched = (decorators & C2_MISMATCHED) != 0; 639 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0; 640 bool on_heap = (decorators & IN_HEAP) != 0; 641 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0; 642 bool is_unordered = (decorators & MO_UNORDERED) != 0; 643 bool need_cpu_mem_bar = !is_unordered || mismatched || !on_heap; 644 645 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : kit->top(); 646 Node* load = BarrierSetC2::load_at_resolved(access, val_type); 647 648 // If we are reading the value of the referent field of a Reference 649 // object (either by using Unsafe directly or through reflection) 650 // then, if SATB is enabled, we need to record the referent in an 651 // SATB log buffer using the pre-barrier mechanism. 652 // Also we need to add memory barrier to prevent commoning reads 653 // from this field across safepoint since GC can change its value. 654 bool need_read_barrier = ShenandoahKeepAliveBarrier && 655 (on_heap && (on_weak || (unknown && offset != kit->top() && obj != kit->top()))); 656 657 if (!access.is_oop() || !need_read_barrier) { 658 return load; 659 } 660 661 if (on_weak) { 662 // Use the pre-barrier to record the value in the referent field 663 satb_write_barrier_pre(kit, false /* do_load */, 664 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 665 load /* pre_val */, T_OBJECT); 666 // Add memory barrier to prevent commoning reads from this field 667 // across safepoint since GC can change its value. 668 kit->insert_mem_bar(Op_MemBarCPUOrder); 669 } else if (unknown) { 670 // We do not require a mem bar inside pre_barrier if need_mem_bar 671 // is set: the barriers would be emitted by us. 672 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar); 673 } 674 675 return load; 676 } 677 678 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicAccess& access, Node* expected_val, 679 Node* new_val, const Type* value_type) const { 680 GraphKit* kit = access.kit(); 681 if (access.is_oop()) { 682 new_val = shenandoah_storeval_barrier(kit, new_val); 683 shenandoah_write_barrier_pre(kit, false /* do_load */, 684 NULL, NULL, max_juint, NULL, NULL, 685 expected_val /* pre_val */, T_OBJECT); 686 687 MemNode::MemOrd mo = access.mem_node_mo(); 688 Node* mem = access.memory(); 689 Node* adr = access.addr().node(); 690 const TypePtr* adr_type = access.addr().type(); 691 Node* load_store = NULL; 692 693 #ifdef _LP64 694 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 695 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop())); 696 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop())); 697 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo)); 698 } else 699 #endif 700 { 701 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo)); 702 } 703 704 access.set_raw_access(load_store); 705 pin_atomic_op(access); 706 707 #ifdef _LP64 708 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 709 return kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 710 } 711 #endif 712 return load_store; 713 } 714 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type); 715 } 716 717 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicAccess& access, Node* expected_val, 718 Node* new_val, const Type* value_type) const { 719 GraphKit* kit = access.kit(); 720 if (access.is_oop()) { 721 new_val = shenandoah_storeval_barrier(kit, new_val); 722 shenandoah_write_barrier_pre(kit, false /* do_load */, 723 NULL, NULL, max_juint, NULL, NULL, 724 expected_val /* pre_val */, T_OBJECT); 725 DecoratorSet decorators = access.decorators(); 726 MemNode::MemOrd mo = access.mem_node_mo(); 727 Node* mem = access.memory(); 728 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0; 729 Node* load_store = NULL; 730 Node* adr = access.addr().node(); 731 #ifdef _LP64 732 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 733 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop())); 734 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop())); 735 if (is_weak_cas) { 736 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 737 } else { 738 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 739 } 740 } else 741 #endif 742 { 743 if (is_weak_cas) { 744 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 745 } else { 746 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 747 } 748 } 749 access.set_raw_access(load_store); 750 pin_atomic_op(access); 751 return load_store; 752 } 753 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type); 754 } 755 756 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicAccess& access, Node* val, const Type* value_type) const { 757 GraphKit* kit = access.kit(); 758 if (access.is_oop()) { 759 val = shenandoah_storeval_barrier(kit, val); 760 } 761 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type); 762 if (access.is_oop()) { 763 shenandoah_write_barrier_pre(kit, false /* do_load */, 764 NULL, NULL, max_juint, NULL, NULL, 765 result /* pre_val */, T_OBJECT); 766 } 767 return result; 768 } 769 770 void ShenandoahBarrierSetC2::clone(GraphKit* kit, Node* src, Node* dst, Node* size, bool is_array) const { 771 assert(!src->is_AddP(), "unexpected input"); 772 src = shenandoah_read_barrier(kit, src); 773 BarrierSetC2::clone(kit, src, dst, size, is_array); 774 } 775 776 Node* ShenandoahBarrierSetC2::resolve(GraphKit* kit, Node* n, DecoratorSet decorators) const { 777 bool is_write = decorators & ACCESS_WRITE; 778 if (is_write) { 779 return shenandoah_write_barrier(kit, n); 780 } else { 781 return shenandoah_read_barrier(kit, n); 782 } 783 } 784 785 Node* ShenandoahBarrierSetC2::obj_allocate(PhaseMacroExpand* macro, Node* ctrl, Node* mem, Node* toobig_false, Node* size_in_bytes, 786 Node*& i_o, Node*& needgc_ctrl, 787 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem, 788 intx prefetch_lines) const { 789 PhaseIterGVN& igvn = macro->igvn(); 790 791 // Allocate several words more for the Shenandoah brooks pointer. 792 size_in_bytes = new AddXNode(size_in_bytes, igvn.MakeConX(BrooksPointer::byte_size())); 793 macro->transform_later(size_in_bytes); 794 795 Node* fast_oop = BarrierSetC2::obj_allocate(macro, ctrl, mem, toobig_false, size_in_bytes, 796 i_o, needgc_ctrl, fast_oop_ctrl, fast_oop_rawmem, 797 prefetch_lines); 798 799 // Bump up object for Shenandoah brooks pointer. 800 fast_oop = new AddPNode(macro->top(), fast_oop, igvn.MakeConX(BrooksPointer::byte_size())); 801 macro->transform_later(fast_oop); 802 803 // Initialize Shenandoah brooks pointer to point to the object itself. 804 fast_oop_rawmem = macro->make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, BrooksPointer::byte_offset(), fast_oop, T_OBJECT); 805 806 return fast_oop; 807 } 808 809 // Support for GC barriers emitted during parsing 810 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const { 811 if (node->Opcode() != Op_CallLeaf) { 812 return false; 813 } 814 CallLeafNode *call = node->as_CallLeaf(); 815 if (call->_name == NULL) { 816 return false; 817 } 818 819 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 || 820 strcmp(call->_name, "shenandoah_cas_obj") == 0 || 821 strcmp(call->_name, "shenandoah_wb_pre") == 0; 822 } 823 824 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const { 825 return ShenandoahBarrierNode::skip_through_barrier(c); 826 } 827 828 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, ArrayCopyPhase phase) const { 829 bool is_oop = type == T_OBJECT || type == T_ARRAY; 830 if (!is_oop) { 831 return false; 832 } 833 834 if (tightly_coupled_alloc) { 835 if (phase == Optimization) { 836 return false; 837 } 838 return !is_clone; 839 } 840 if (phase == Optimization) { 841 return !ShenandoahStoreValEnqueueBarrier; 842 } 843 return true; 844 } 845 846 Node* ShenandoahBarrierSetC2::array_copy_load_store_barrier(PhaseGVN *phase, bool can_reshape, Node* v, MergeMemNode* mem, Node*& ctl) const { 847 if (ShenandoahStoreValReadBarrier) { 848 RegionNode* region = new RegionNode(3); 849 const Type* v_t = phase->type(v); 850 Node* phi = new PhiNode(region, v_t); 851 Node* cmp = phase->transform(new CmpPNode(v, phase->zerocon(T_OBJECT))); 852 Node* bol = phase->transform(new BoolNode(cmp, BoolTest::ne)); 853 IfNode* iff = new IfNode(ctl, bol, PROB_LIKELY_MAG(3), COUNT_UNKNOWN); 854 855 phase->transform(iff); 856 if (can_reshape) { 857 phase->is_IterGVN()->_worklist.push(iff); 858 } else { 859 phase->record_for_igvn(iff); 860 } 861 862 Node* null_true = phase->transform(new IfFalseNode(iff)); 863 Node* null_false = phase->transform(new IfTrueNode(iff)); 864 region->init_req(1, null_true); 865 region->init_req(2, null_false); 866 phi->init_req(1, phase->zerocon(T_OBJECT)); 867 Node* cast = new CastPPNode(v, phase->type(v)->join_speculative(TypePtr::NOTNULL)); 868 cast->set_req(0, null_false); 869 cast = phase->transform(cast); 870 Node* rb = phase->transform(new ShenandoahReadBarrierNode(null_false, phase->C->immutable_memory(), cast, false)); 871 phi->init_req(2, rb); 872 ctl = phase->transform(region); 873 return phase->transform(phi); 874 } 875 if (ShenandoahStoreValEnqueueBarrier) { 876 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(phase->type(v)); 877 int alias = phase->C->get_alias_index(adr_type); 878 Node* wb = new ShenandoahWriteBarrierNode(phase->C, ctl, mem->memory_at(alias), v); 879 Node* wb_transformed = phase->transform(wb); 880 Node* enqueue = phase->transform(new ShenandoahEnqueueBarrierNode(wb_transformed)); 881 if (wb_transformed == wb) { 882 Node* proj = phase->transform(new ShenandoahWBMemProjNode(wb)); 883 mem->set_memory_at(alias, proj); 884 } 885 return enqueue; 886 } 887 return v; 888 } 889 890 void ShenandoahBarrierSetC2::array_copy_post_barrier_at_expansion(ArrayCopyNode* ac, Node*& c, Node*& m, PhaseIterGVN& igvn) const { 891 assert(ac->is_clonebasic(), "no other kind of arraycopy here"); 892 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 893 Node* dest = ac->in(ArrayCopyNode::Dest); 894 assert(dest->is_AddP(), "bad input"); 895 Node* call = new CallLeafNoFPNode(ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(), 896 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier), 897 "shenandoah_clone_barrier", raw_adr_type); 898 call->init_req(TypeFunc::Control, c); 899 call->init_req(TypeFunc::I_O , igvn.C->top()); 900 call->init_req(TypeFunc::Memory , m); 901 call->init_req(TypeFunc::ReturnAdr, igvn.C->top()); 902 call->init_req(TypeFunc::FramePtr, igvn.C->top()); 903 call->init_req(TypeFunc::Parms+0, dest->in(AddPNode::Base)); 904 905 call = igvn.transform(call); 906 c = new ProjNode(call,TypeFunc::Control); 907 c = igvn.transform(c); 908 m = new ProjNode(call, TypeFunc::Memory); 909 m = igvn.transform(m); 910 } 911 912 913 // Support for macro expanded GC barriers 914 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const { 915 if (node->Opcode() == Op_ShenandoahWriteBarrier) { 916 state()->add_shenandoah_barrier((ShenandoahWriteBarrierNode*) node); 917 } 918 } 919 920 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const { 921 if (node->Opcode() == Op_ShenandoahWriteBarrier) { 922 state()->remove_shenandoah_barrier((ShenandoahWriteBarrierNode*) node); 923 } 924 } 925 926 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const { 927 if (is_shenandoah_wb_pre_call(n)) { 928 shenandoah_eliminate_wb_pre(n, ¯o->igvn()); 929 } 930 } 931 932 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const { 933 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), ""); 934 Node* c = call->as_Call()->proj_out(TypeFunc::Control); 935 c = c->unique_ctrl_out(); 936 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 937 c = c->unique_ctrl_out(); 938 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 939 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0); 940 assert(iff->is_If(), "expect test"); 941 if (!is_shenandoah_marking_if(igvn, iff)) { 942 c = c->unique_ctrl_out(); 943 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 944 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0); 945 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test"); 946 } 947 Node* cmpx = iff->in(1)->in(1); 948 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ)); 949 igvn->rehash_node_delayed(call); 950 call->del_req(call->req()-1); 951 } 952 953 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const { 954 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) { 955 igvn->add_users_to_worklist(node); 956 } 957 } 958 959 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const { 960 for (uint i = 0; i < useful.size(); i++) { 961 Node* n = useful.at(i); 962 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) { 963 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 964 C->record_for_igvn(n->fast_out(i)); 965 } 966 } 967 } 968 for (int i = state()->shenandoah_barriers_count()-1; i >= 0; i--) { 969 ShenandoahWriteBarrierNode* n = state()->shenandoah_barrier(i); 970 if (!useful.member(n)) { 971 state()->remove_shenandoah_barrier(n); 972 } 973 } 974 975 } 976 977 void ShenandoahBarrierSetC2::add_users_to_worklist(Unique_Node_List* worklist) const {} 978 979 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const { 980 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena); 981 } 982 983 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const { 984 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state()); 985 } 986 987 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be 988 // expanded later, then now is the time to do so. 989 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; } 990 void ShenandoahBarrierSetC2::verify_gc_barriers(bool post_parse) const { 991 #ifdef ASSERT 992 if (ShenandoahVerifyOptoBarriers && !post_parse) { 993 ShenandoahBarrierNode::verify(Compile::current()->root()); 994 } 995 #endif 996 } 997 998 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN *phase, Node* n, bool can_reshape) const { 999 if (is_shenandoah_wb_pre_call(n)) { 1000 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt(); 1001 if (n->req() > cnt) { 1002 Node* addp = n->in(cnt); 1003 if (has_only_shenandoah_wb_pre_uses(addp)) { 1004 n->del_req(cnt); 1005 if (can_reshape) { 1006 phase->is_IterGVN()->_worklist.push(addp); 1007 } 1008 return n; 1009 } 1010 } 1011 } 1012 if (n->Opcode() == Op_CmpP) { 1013 Node* in1 = n->in(1); 1014 Node* in2 = n->in(2); 1015 if (in1->bottom_type() == TypePtr::NULL_PTR) { 1016 in2 = step_over_gc_barrier(in2); 1017 } 1018 if (in2->bottom_type() == TypePtr::NULL_PTR) { 1019 in1 = step_over_gc_barrier(in1); 1020 } 1021 PhaseIterGVN* igvn = phase->is_IterGVN(); 1022 if (in1 != n->in(1)) { 1023 if (igvn != NULL) { 1024 n->set_req_X(1, in1, igvn); 1025 } else { 1026 n->set_req(1, in1); 1027 } 1028 assert(in2 == n->in(2), "only one change"); 1029 return n; 1030 } 1031 if (in2 != n->in(2)) { 1032 if (igvn != NULL) { 1033 n->set_req_X(2, in2, igvn); 1034 } else { 1035 n->set_req(2, in2); 1036 } 1037 return n; 1038 } 1039 } 1040 1041 return NULL; 1042 } 1043 1044 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) { 1045 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1046 Node* u = n->fast_out(i); 1047 if (!is_shenandoah_wb_pre_call(u)) { 1048 return false; 1049 } 1050 } 1051 return n->outcnt() > 0; 1052 }