1 /* 2 * Copyright (c) 2018, Red Hat, Inc. All rights reserved. 3 * 4 * This code is free software; you can redistribute it and/or modify it 5 * under the terms of the GNU General Public License version 2 only, as 6 * published by the Free Software Foundation. 7 * 8 * This code is distributed in the hope that it will be useful, but WITHOUT 9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 11 * version 2 for more details (a copy is included in the LICENSE file that 12 * accompanied this code). 13 * 14 * You should have received a copy of the GNU General Public License version 15 * 2 along with this work; if not, write to the Free Software Foundation, 16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 17 * 18 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 19 * or visit www.oracle.com if you need additional information or have any 20 * questions. 21 * 22 */ 23 24 #include "precompiled.hpp" 25 #include "gc/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/escape.hpp" 32 #include "opto/graphKit.hpp" 33 #include "opto/idealKit.hpp" 34 #include "opto/macro.hpp" 35 #include "opto/movenode.hpp" 36 #include "opto/narrowptrnode.hpp" 37 #include "opto/rootnode.hpp" 38 39 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() { 40 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2()); 41 } 42 43 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena) 44 : _shenandoah_barriers(new (comp_arena) GrowableArray<ShenandoahWriteBarrierNode*>(comp_arena, 8, 0, NULL)) { 45 } 46 47 int ShenandoahBarrierSetC2State::shenandoah_barriers_count() const { 48 return _shenandoah_barriers->length(); 49 } 50 51 ShenandoahWriteBarrierNode* ShenandoahBarrierSetC2State::shenandoah_barrier(int idx) const { 52 return _shenandoah_barriers->at(idx); 53 } 54 55 void ShenandoahBarrierSetC2State::add_shenandoah_barrier(ShenandoahWriteBarrierNode * n) { 56 assert(!_shenandoah_barriers->contains(n), "duplicate entry in barrier list"); 57 _shenandoah_barriers->append(n); 58 } 59 60 void ShenandoahBarrierSetC2State::remove_shenandoah_barrier(ShenandoahWriteBarrierNode * n) { 61 if (_shenandoah_barriers->contains(n)) { 62 _shenandoah_barriers->remove(n); 63 } 64 } 65 66 #define __ kit-> 67 68 Node* ShenandoahBarrierSetC2::shenandoah_read_barrier(GraphKit* kit, Node* obj) const { 69 if (ShenandoahReadBarrier) { 70 obj = shenandoah_read_barrier_impl(kit, obj, false, true, true); 71 } 72 return obj; 73 } 74 75 Node* ShenandoahBarrierSetC2::shenandoah_storeval_barrier(GraphKit* kit, Node* obj) const { 76 if (ShenandoahStoreValEnqueueBarrier) { 77 obj = shenandoah_write_barrier(kit, obj); 78 obj = shenandoah_enqueue_barrier(kit, obj); 79 } 80 if (ShenandoahStoreValReadBarrier) { 81 obj = shenandoah_read_barrier_impl(kit, obj, true, false, false); 82 } 83 return obj; 84 } 85 86 Node* ShenandoahBarrierSetC2::shenandoah_read_barrier_acmp(GraphKit* kit, Node* obj) { 87 return shenandoah_read_barrier_impl(kit, obj, true, true, false); 88 } 89 90 Node* ShenandoahBarrierSetC2::shenandoah_read_barrier_impl(GraphKit* kit, Node* obj, bool use_ctrl, bool use_mem, bool allow_fromspace) const { 91 const Type* obj_type = obj->bottom_type(); 92 if (obj_type->higher_equal(TypePtr::NULL_PTR)) { 93 return obj; 94 } 95 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type); 96 Node* mem = use_mem ? __ memory(adr_type) : __ immutable_memory(); 97 98 if (! ShenandoahBarrierNode::needs_barrier(&__ gvn(), NULL, obj, mem, allow_fromspace)) { 99 // We know it is null, no barrier needed. 100 return obj; 101 } 102 103 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) { 104 105 // We don't know if it's null or not. Need null-check. 106 enum { _not_null_path = 1, _null_path, PATH_LIMIT }; 107 RegionNode* region = new RegionNode(PATH_LIMIT); 108 Node* phi = new PhiNode(region, obj_type); 109 Node* null_ctrl = __ top(); 110 Node* not_null_obj = __ null_check_oop(obj, &null_ctrl); 111 112 region->init_req(_null_path, null_ctrl); 113 phi ->init_req(_null_path, __ zerocon(T_OBJECT)); 114 115 Node* ctrl = use_ctrl ? __ control() : NULL; 116 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, not_null_obj, allow_fromspace); 117 Node* n = __ gvn().transform(rb); 118 119 region->init_req(_not_null_path, __ control()); 120 phi ->init_req(_not_null_path, n); 121 122 __ set_control(__ gvn().transform(region)); 123 __ record_for_igvn(region); 124 return __ gvn().transform(phi); 125 126 } else { 127 // We know it is not null. Simple barrier is sufficient. 128 Node* ctrl = use_ctrl ? __ control() : NULL; 129 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, obj, allow_fromspace); 130 Node* n = __ gvn().transform(rb); 131 __ record_for_igvn(n); 132 return n; 133 } 134 } 135 136 Node* ShenandoahBarrierSetC2::shenandoah_write_barrier_helper(GraphKit* kit, Node* obj, const TypePtr* adr_type) const { 137 ShenandoahWriteBarrierNode* wb = new ShenandoahWriteBarrierNode(kit->C, kit->control(), kit->memory(adr_type), obj); 138 Node* n = __ gvn().transform(wb); 139 if (n == wb) { // New barrier needs memory projection. 140 Node* proj = __ gvn().transform(new ShenandoahWBMemProjNode(n)); 141 __ set_memory(proj, adr_type); 142 } 143 return n; 144 } 145 146 Node* ShenandoahBarrierSetC2::shenandoah_write_barrier(GraphKit* kit, Node* obj) const { 147 if (ShenandoahWriteBarrier) { 148 obj = shenandoah_write_barrier_impl(kit, obj); 149 } 150 return obj; 151 } 152 153 Node* ShenandoahBarrierSetC2::shenandoah_write_barrier_impl(GraphKit* kit, Node* obj) const { 154 if (! ShenandoahBarrierNode::needs_barrier(&__ gvn(), NULL, obj, NULL, true)) { 155 return obj; 156 } 157 const Type* obj_type = obj->bottom_type(); 158 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type); 159 Node* n = shenandoah_write_barrier_helper(kit, obj, adr_type); 160 __ record_for_igvn(n); 161 return n; 162 } 163 164 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseTransform* phase, Node* adr, 165 BasicType bt, uint adr_idx) const { 166 intptr_t offset = 0; 167 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 168 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 169 170 if (offset == Type::OffsetBot) { 171 return false; // cannot unalias unless there are precise offsets 172 } 173 174 if (alloc == NULL) { 175 return false; // No allocation found 176 } 177 178 intptr_t size_in_bytes = type2aelembytes(bt); 179 180 Node* mem = __ memory(adr_idx); // start searching here... 181 182 for (int cnt = 0; cnt < 50; cnt++) { 183 184 if (mem->is_Store()) { 185 186 Node* st_adr = mem->in(MemNode::Address); 187 intptr_t st_offset = 0; 188 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 189 190 if (st_base == NULL) { 191 break; // inscrutable pointer 192 } 193 194 // Break we have found a store with same base and offset as ours so break 195 if (st_base == base && st_offset == offset) { 196 break; 197 } 198 199 if (st_offset != offset && st_offset != Type::OffsetBot) { 200 const int MAX_STORE = BytesPerLong; 201 if (st_offset >= offset + size_in_bytes || 202 st_offset <= offset - MAX_STORE || 203 st_offset <= offset - mem->as_Store()->memory_size()) { 204 // Success: The offsets are provably independent. 205 // (You may ask, why not just test st_offset != offset and be done? 206 // The answer is that stores of different sizes can co-exist 207 // in the same sequence of RawMem effects. We sometimes initialize 208 // a whole 'tile' of array elements with a single jint or jlong.) 209 mem = mem->in(MemNode::Memory); 210 continue; // advance through independent store memory 211 } 212 } 213 214 if (st_base != base 215 && MemNode::detect_ptr_independence(base, alloc, st_base, 216 AllocateNode::Ideal_allocation(st_base, phase), 217 phase)) { 218 // Success: The bases are provably independent. 219 mem = mem->in(MemNode::Memory); 220 continue; // advance through independent store memory 221 } 222 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 223 224 InitializeNode* st_init = mem->in(0)->as_Initialize(); 225 AllocateNode* st_alloc = st_init->allocation(); 226 227 // Make sure that we are looking at the same allocation site. 228 // The alloc variable is guaranteed to not be null here from earlier check. 229 if (alloc == st_alloc) { 230 // Check that the initialization is storing NULL so that no previous store 231 // has been moved up and directly write a reference 232 Node* captured_store = st_init->find_captured_store(offset, 233 type2aelembytes(T_OBJECT), 234 phase); 235 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 236 return true; 237 } 238 } 239 } 240 241 // Unless there is an explicit 'continue', we must bail out here, 242 // because 'mem' is an inscrutable memory state (e.g., a call). 243 break; 244 } 245 246 return false; 247 } 248 249 #undef __ 250 #define __ ideal. 251 252 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit, 253 bool do_load, 254 Node* obj, 255 Node* adr, 256 uint alias_idx, 257 Node* val, 258 const TypeOopPtr* val_type, 259 Node* pre_val, 260 BasicType bt) const { 261 // Some sanity checks 262 // Note: val is unused in this routine. 263 264 if (do_load) { 265 // We need to generate the load of the previous value 266 assert(obj != NULL, "must have a base"); 267 assert(adr != NULL, "where are loading from?"); 268 assert(pre_val == NULL, "loaded already?"); 269 assert(val_type != NULL, "need a type"); 270 271 if (ReduceInitialCardMarks 272 && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) { 273 return; 274 } 275 276 } else { 277 // In this case both val_type and alias_idx are unused. 278 assert(pre_val != NULL, "must be loaded already"); 279 // Nothing to be done if pre_val is null. 280 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 281 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 282 } 283 assert(bt == T_OBJECT, "or we shouldn't be here"); 284 285 IdealKit ideal(kit, true); 286 287 Node* tls = __ thread(); // ThreadLocalStorage 288 289 Node* no_base = __ top(); 290 Node* zero = __ ConI(0); 291 Node* zeroX = __ ConX(0); 292 293 float likely = PROB_LIKELY(0.999); 294 float unlikely = PROB_UNLIKELY(0.999); 295 296 // Offsets into the thread 297 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()); 298 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()); 299 300 // Now the actual pointers into the thread 301 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 302 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 303 304 // Now some of the values 305 Node* marking; 306 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()))); 307 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw); 308 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING)); 309 assert(ShenandoahWriteBarrierNode::is_gc_state_load(ld), "Should match the shape"); 310 311 // if (!marking) 312 __ if_then(marking, BoolTest::ne, zero, unlikely); { 313 BasicType index_bt = TypeX_X->basic_type(); 314 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 315 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 316 317 if (do_load) { 318 // load original value 319 // alias_idx correct?? 320 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 321 } 322 323 // if (pre_val != NULL) 324 __ if_then(pre_val, BoolTest::ne, kit->null()); { 325 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 326 327 // is the queue for this thread full? 328 __ if_then(index, BoolTest::ne, zeroX, likely); { 329 330 // decrement the index 331 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 332 333 // Now get the buffer location we will log the previous value into and store it 334 Node *log_addr = __ AddP(no_base, buffer, next_index); 335 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 336 // update the index 337 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 338 339 } __ else_(); { 340 341 // logging buffer is full, call the runtime 342 const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type(); 343 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls); 344 } __ end_if(); // (!index) 345 } __ end_if(); // (pre_val != NULL) 346 } __ end_if(); // (!marking) 347 348 // Final sync IdealKit and GraphKit. 349 kit->final_sync(ideal); 350 351 if (ShenandoahSATBBarrier && adr != NULL) { 352 Node* c = kit->control(); 353 Node* call = c->in(1)->in(1)->in(1)->in(0); 354 assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected"); 355 call->add_req(adr); 356 } 357 } 358 359 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) { 360 return call->is_CallLeaf() && 361 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry); 362 } 363 364 bool ShenandoahBarrierSetC2::is_shenandoah_wb_call(Node* call) { 365 return call->is_CallLeaf() && 366 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_JRT); 367 } 368 369 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseTransform *phase, Node* n) { 370 if (n->Opcode() != Op_If) { 371 return false; 372 } 373 374 Node* bol = n->in(1); 375 assert(bol->is_Bool(), ""); 376 Node* cmpx = bol->in(1); 377 if (bol->as_Bool()->_test._test == BoolTest::ne && 378 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) && 379 is_shenandoah_state_load(cmpx->in(1)->in(1)) && 380 cmpx->in(1)->in(2)->is_Con() && 381 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) { 382 return true; 383 } 384 385 return false; 386 } 387 388 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) { 389 if (!n->is_Load()) return false; 390 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset()); 391 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal 392 && n->in(2)->in(3)->is_Con() 393 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset; 394 } 395 396 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit, 397 bool do_load, 398 Node* obj, 399 Node* adr, 400 uint alias_idx, 401 Node* val, 402 const TypeOopPtr* val_type, 403 Node* pre_val, 404 BasicType bt) const { 405 if (ShenandoahSATBBarrier) { 406 IdealKit ideal(kit); 407 kit->sync_kit(ideal); 408 409 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt); 410 411 ideal.sync_kit(kit); 412 kit->final_sync(ideal); 413 } 414 } 415 416 Node* ShenandoahBarrierSetC2::shenandoah_enqueue_barrier(GraphKit* kit, Node* pre_val) const { 417 return kit->gvn().transform(new ShenandoahEnqueueBarrierNode(pre_val)); 418 } 419 420 // Helper that guards and inserts a pre-barrier. 421 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset, 422 Node* pre_val, bool need_mem_bar) const { 423 // We could be accessing the referent field of a reference object. If so, when G1 424 // is enabled, we need to log the value in the referent field in an SATB buffer. 425 // This routine performs some compile time filters and generates suitable 426 // runtime filters that guard the pre-barrier code. 427 // Also add memory barrier for non volatile load from the referent field 428 // to prevent commoning of loads across safepoint. 429 430 // Some compile time checks. 431 432 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 433 const TypeX* otype = offset->find_intptr_t_type(); 434 if (otype != NULL && otype->is_con() && 435 otype->get_con() != java_lang_ref_Reference::referent_offset) { 436 // Constant offset but not the reference_offset so just return 437 return; 438 } 439 440 // We only need to generate the runtime guards for instances. 441 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 442 if (btype != NULL) { 443 if (btype->isa_aryptr()) { 444 // Array type so nothing to do 445 return; 446 } 447 448 const TypeInstPtr* itype = btype->isa_instptr(); 449 if (itype != NULL) { 450 // Can the klass of base_oop be statically determined to be 451 // _not_ a sub-class of Reference and _not_ Object? 452 ciKlass* klass = itype->klass(); 453 if ( klass->is_loaded() && 454 !klass->is_subtype_of(kit->env()->Reference_klass()) && 455 !kit->env()->Object_klass()->is_subtype_of(klass)) { 456 return; 457 } 458 } 459 } 460 461 // The compile time filters did not reject base_oop/offset so 462 // we need to generate the following runtime filters 463 // 464 // if (offset == java_lang_ref_Reference::_reference_offset) { 465 // if (instance_of(base, java.lang.ref.Reference)) { 466 // pre_barrier(_, pre_val, ...); 467 // } 468 // } 469 470 float likely = PROB_LIKELY( 0.999); 471 float unlikely = PROB_UNLIKELY(0.999); 472 473 IdealKit ideal(kit); 474 475 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 476 477 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 478 // Update graphKit memory and control from IdealKit. 479 kit->sync_kit(ideal); 480 481 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass())); 482 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con); 483 484 // Update IdealKit memory and control from graphKit. 485 __ sync_kit(kit); 486 487 Node* one = __ ConI(1); 488 // is_instof == 0 if base_oop == NULL 489 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 490 491 // Update graphKit from IdeakKit. 492 kit->sync_kit(ideal); 493 494 // Use the pre-barrier to record the value in the referent field 495 satb_write_barrier_pre(kit, false /* do_load */, 496 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 497 pre_val /* pre_val */, 498 T_OBJECT); 499 if (need_mem_bar) { 500 // Add memory barrier to prevent commoning reads from this field 501 // across safepoint since GC can change its value. 502 kit->insert_mem_bar(Op_MemBarCPUOrder); 503 } 504 // Update IdealKit from graphKit. 505 __ sync_kit(kit); 506 507 } __ end_if(); // _ref_type != ref_none 508 } __ end_if(); // offset == referent_offset 509 510 // Final sync IdealKit and GraphKit. 511 kit->final_sync(ideal); 512 } 513 514 #undef __ 515 516 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() { 517 const Type **fields = TypeTuple::fields(2); 518 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value 519 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread 520 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields); 521 522 // create result type (range) 523 fields = TypeTuple::fields(0); 524 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields); 525 526 return TypeFunc::make(domain, range); 527 } 528 529 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() { 530 const Type **fields = TypeTuple::fields(1); 531 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value 532 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields); 533 534 // create result type (range) 535 fields = TypeTuple::fields(0); 536 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields); 537 538 return TypeFunc::make(domain, range); 539 } 540 541 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_write_barrier_Type() { 542 const Type **fields = TypeTuple::fields(1); 543 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value 544 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields); 545 546 // create result type (range) 547 fields = TypeTuple::fields(1); 548 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; 549 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields); 550 551 return TypeFunc::make(domain, range); 552 } 553 554 void ShenandoahBarrierSetC2::resolve_address(C2Access& access) const { 555 const TypePtr* adr_type = access.addr().type(); 556 557 if ((access.decorators() & IN_NATIVE) == 0 && (adr_type->isa_instptr() || adr_type->isa_aryptr())) { 558 int off = adr_type->is_ptr()->offset(); 559 int base_off = adr_type->isa_instptr() ? instanceOopDesc::base_offset_in_bytes() : 560 arrayOopDesc::base_offset_in_bytes(adr_type->is_aryptr()->elem()->array_element_basic_type()); 561 assert(off != Type::OffsetTop, "unexpected offset"); 562 if (off == Type::OffsetBot || off >= base_off) { 563 DecoratorSet decorators = access.decorators(); 564 bool is_write = (decorators & C2_WRITE_ACCESS) != 0; 565 GraphKit* kit = NULL; 566 if (access.is_parse_access()) { 567 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access); 568 kit = parse_access.kit(); 569 } 570 Node* adr = access.addr().node(); 571 assert(adr->is_AddP(), "unexpected address shape"); 572 Node* base = adr->in(AddPNode::Base); 573 574 if (is_write) { 575 if (kit != NULL) { 576 base = shenandoah_write_barrier(kit, base); 577 } else { 578 assert(access.is_opt_access(), "either parse or opt access"); 579 assert((access.decorators() & C2_ARRAY_COPY) != 0, "can be skipped for clone"); 580 } 581 } else { 582 if (adr_type->isa_instptr()) { 583 Compile* C = access.gvn().C; 584 ciField* field = C->alias_type(adr_type)->field(); 585 586 // Insert read barrier for Shenandoah. 587 if (field != NULL && 588 ((ShenandoahOptimizeStaticFinals && field->is_static() && field->is_final()) || 589 (ShenandoahOptimizeInstanceFinals && !field->is_static() && field->is_final()) || 590 (ShenandoahOptimizeStableFinals && field->is_stable()))) { 591 // Skip the barrier for special fields 592 } else { 593 if (kit != NULL) { 594 base = shenandoah_read_barrier(kit, base); 595 } else { 596 assert(access.is_opt_access(), "either parse or opt access"); 597 assert((access.decorators() & C2_ARRAY_COPY) != 0, "can be skipped for arraycopy"); 598 } 599 } 600 } else { 601 if (kit != NULL) { 602 base = shenandoah_read_barrier(kit, base); 603 } else { 604 assert(access.is_opt_access(), "either parse or opt access"); 605 assert((access.decorators() & C2_ARRAY_COPY) != 0, "can be skipped for arraycopy"); 606 } 607 } 608 } 609 if (base != adr->in(AddPNode::Base)) { 610 assert(kit != NULL, "no barrier should have been added"); 611 612 Node* address = adr->in(AddPNode::Address); 613 614 if (address->is_AddP()) { 615 assert(address->in(AddPNode::Base) == adr->in(AddPNode::Base), "unexpected address shape"); 616 assert(!address->in(AddPNode::Address)->is_AddP(), "unexpected address shape"); 617 assert(address->in(AddPNode::Address) == adr->in(AddPNode::Base), "unexpected address shape"); 618 address = address->clone(); 619 address->set_req(AddPNode::Base, base); 620 address->set_req(AddPNode::Address, base); 621 address = kit->gvn().transform(address); 622 } else { 623 assert(address == adr->in(AddPNode::Base), "unexpected address shape"); 624 address = base; 625 } 626 adr = adr->clone(); 627 adr->set_req(AddPNode::Base, base); 628 adr->set_req(AddPNode::Address, address); 629 adr = kit->gvn().transform(adr); 630 access.addr().set_node(adr); 631 } 632 } 633 } 634 } 635 636 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const { 637 DecoratorSet decorators = access.decorators(); 638 639 const TypePtr* adr_type = access.addr().type(); 640 Node* adr = access.addr().node(); 641 642 bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0; 643 bool on_heap = (decorators & IN_HEAP) != 0; 644 645 if (!access.is_oop() || (!on_heap && !anonymous)) { 646 return BarrierSetC2::store_at_resolved(access, val); 647 } 648 649 if (access.is_parse_access()) { 650 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access); 651 GraphKit* kit = parse_access.kit(); 652 653 uint adr_idx = kit->C->get_alias_index(adr_type); 654 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 655 Node* value = val.node(); 656 value = shenandoah_storeval_barrier(kit, value); 657 val.set_node(value); 658 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(), 659 static_cast<const TypeOopPtr*>(val.type()), NULL /* pre_val */, access.type()); 660 } else { 661 assert(access.is_opt_access(), "only for optimization passes"); 662 assert(((decorators & C2_TIGHLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code"); 663 C2OptAccess& opt_access = static_cast<C2OptAccess&>(access); 664 PhaseGVN& gvn = opt_access.gvn(); 665 MergeMemNode* mm = opt_access.mem(); 666 667 if (ShenandoahStoreValReadBarrier) { 668 RegionNode* region = new RegionNode(3); 669 const Type* v_t = gvn.type(val.node()); 670 Node* phi = new PhiNode(region, v_t->isa_oopptr() ? v_t->is_oopptr()->cast_to_nonconst() : v_t); 671 Node* cmp = gvn.transform(new CmpPNode(val.node(), gvn.zerocon(T_OBJECT))); 672 Node* bol = gvn.transform(new BoolNode(cmp, BoolTest::ne)); 673 IfNode* iff = new IfNode(opt_access.ctl(), bol, PROB_LIKELY_MAG(3), COUNT_UNKNOWN); 674 675 gvn.transform(iff); 676 if (gvn.is_IterGVN()) { 677 gvn.is_IterGVN()->_worklist.push(iff); 678 } else { 679 gvn.record_for_igvn(iff); 680 } 681 682 Node* null_true = gvn.transform(new IfFalseNode(iff)); 683 Node* null_false = gvn.transform(new IfTrueNode(iff)); 684 region->init_req(1, null_true); 685 region->init_req(2, null_false); 686 phi->init_req(1, gvn.zerocon(T_OBJECT)); 687 Node* cast = new CastPPNode(val.node(), gvn.type(val.node())->join_speculative(TypePtr::NOTNULL)); 688 cast->set_req(0, null_false); 689 cast = gvn.transform(cast); 690 Node* rb = gvn.transform(new ShenandoahReadBarrierNode(null_false, gvn.C->immutable_memory(), cast, false)); 691 phi->init_req(2, rb); 692 opt_access.set_ctl(gvn.transform(region)); 693 val.set_node(gvn.transform(phi)); 694 } 695 if (ShenandoahStoreValEnqueueBarrier) { 696 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(gvn.type(val.node())); 697 int alias = gvn.C->get_alias_index(adr_type); 698 Node* wb = new ShenandoahWriteBarrierNode(gvn.C, opt_access.ctl(), mm->memory_at(alias), val.node()); 699 Node* wb_transformed = gvn.transform(wb); 700 Node* enqueue = gvn.transform(new ShenandoahEnqueueBarrierNode(wb_transformed)); 701 if (wb_transformed == wb) { 702 Node* proj = gvn.transform(new ShenandoahWBMemProjNode(wb)); 703 mm->set_memory_at(alias, proj); 704 } 705 val.set_node(enqueue); 706 } 707 } 708 return BarrierSetC2::store_at_resolved(access, val); 709 } 710 711 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const { 712 DecoratorSet decorators = access.decorators(); 713 714 Node* adr = access.addr().node(); 715 Node* obj = access.base(); 716 717 bool mismatched = (decorators & C2_MISMATCHED) != 0; 718 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0; 719 bool on_heap = (decorators & IN_HEAP) != 0; 720 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0; 721 bool is_unordered = (decorators & MO_UNORDERED) != 0; 722 bool need_cpu_mem_bar = !is_unordered || mismatched || !on_heap; 723 724 Node* top = Compile::current()->top(); 725 726 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top; 727 Node* load = BarrierSetC2::load_at_resolved(access, val_type); 728 729 // If we are reading the value of the referent field of a Reference 730 // object (either by using Unsafe directly or through reflection) 731 // then, if SATB is enabled, we need to record the referent in an 732 // SATB log buffer using the pre-barrier mechanism. 733 // Also we need to add memory barrier to prevent commoning reads 734 // from this field across safepoint since GC can change its value. 735 bool need_read_barrier = ShenandoahKeepAliveBarrier && 736 (on_heap && (on_weak || (unknown && offset != top && obj != top))); 737 738 if (!access.is_oop() || !need_read_barrier) { 739 return load; 740 } 741 742 assert(access.is_parse_access(), "entry not supported at optimization time"); 743 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access); 744 GraphKit* kit = parse_access.kit(); 745 746 if (on_weak) { 747 // Use the pre-barrier to record the value in the referent field 748 satb_write_barrier_pre(kit, false /* do_load */, 749 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 750 load /* pre_val */, T_OBJECT); 751 // Add memory barrier to prevent commoning reads from this field 752 // across safepoint since GC can change its value. 753 kit->insert_mem_bar(Op_MemBarCPUOrder); 754 } else if (unknown) { 755 // We do not require a mem bar inside pre_barrier if need_mem_bar 756 // is set: the barriers would be emitted by us. 757 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar); 758 } 759 760 return load; 761 } 762 763 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val, 764 Node* new_val, const Type* value_type) const { 765 GraphKit* kit = access.kit(); 766 if (access.is_oop()) { 767 new_val = shenandoah_storeval_barrier(kit, new_val); 768 shenandoah_write_barrier_pre(kit, false /* do_load */, 769 NULL, NULL, max_juint, NULL, NULL, 770 expected_val /* pre_val */, T_OBJECT); 771 772 MemNode::MemOrd mo = access.mem_node_mo(); 773 Node* mem = access.memory(); 774 Node* adr = access.addr().node(); 775 const TypePtr* adr_type = access.addr().type(); 776 Node* load_store = NULL; 777 778 #ifdef _LP64 779 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 780 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop())); 781 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop())); 782 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo)); 783 } else 784 #endif 785 { 786 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo)); 787 } 788 789 access.set_raw_access(load_store); 790 pin_atomic_op(access); 791 792 #ifdef _LP64 793 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 794 return kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 795 } 796 #endif 797 return load_store; 798 } 799 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type); 800 } 801 802 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val, 803 Node* new_val, const Type* value_type) const { 804 GraphKit* kit = access.kit(); 805 if (access.is_oop()) { 806 new_val = shenandoah_storeval_barrier(kit, new_val); 807 shenandoah_write_barrier_pre(kit, false /* do_load */, 808 NULL, NULL, max_juint, NULL, NULL, 809 expected_val /* pre_val */, T_OBJECT); 810 DecoratorSet decorators = access.decorators(); 811 MemNode::MemOrd mo = access.mem_node_mo(); 812 Node* mem = access.memory(); 813 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0; 814 Node* load_store = NULL; 815 Node* adr = access.addr().node(); 816 #ifdef _LP64 817 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 818 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop())); 819 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop())); 820 if (is_weak_cas) { 821 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 822 } else { 823 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo)); 824 } 825 } else 826 #endif 827 { 828 if (is_weak_cas) { 829 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 830 } else { 831 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo)); 832 } 833 } 834 access.set_raw_access(load_store); 835 pin_atomic_op(access); 836 return load_store; 837 } 838 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type); 839 } 840 841 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const { 842 GraphKit* kit = access.kit(); 843 if (access.is_oop()) { 844 val = shenandoah_storeval_barrier(kit, val); 845 } 846 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type); 847 if (access.is_oop()) { 848 shenandoah_write_barrier_pre(kit, false /* do_load */, 849 NULL, NULL, max_juint, NULL, NULL, 850 result /* pre_val */, T_OBJECT); 851 } 852 return result; 853 } 854 855 void ShenandoahBarrierSetC2::clone(GraphKit* kit, Node* src, Node* dst, Node* size, bool is_array) const { 856 assert(!src->is_AddP(), "unexpected input"); 857 src = shenandoah_read_barrier(kit, src); 858 BarrierSetC2::clone(kit, src, dst, size, is_array); 859 } 860 861 Node* ShenandoahBarrierSetC2::resolve(GraphKit* kit, Node* n, DecoratorSet decorators) const { 862 bool is_write = decorators & ACCESS_WRITE; 863 if (is_write) { 864 return shenandoah_write_barrier(kit, n); 865 } else { 866 return shenandoah_read_barrier(kit, n); 867 } 868 } 869 870 Node* ShenandoahBarrierSetC2::obj_allocate(PhaseMacroExpand* macro, Node* ctrl, Node* mem, Node* toobig_false, Node* size_in_bytes, 871 Node*& i_o, Node*& needgc_ctrl, 872 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem, 873 intx prefetch_lines) const { 874 PhaseIterGVN& igvn = macro->igvn(); 875 876 // Allocate several words more for the Shenandoah brooks pointer. 877 size_in_bytes = new AddXNode(size_in_bytes, igvn.MakeConX(ShenandoahBrooksPointer::byte_size())); 878 macro->transform_later(size_in_bytes); 879 880 Node* fast_oop = BarrierSetC2::obj_allocate(macro, ctrl, mem, toobig_false, size_in_bytes, 881 i_o, needgc_ctrl, fast_oop_ctrl, fast_oop_rawmem, 882 prefetch_lines); 883 884 // Bump up object for Shenandoah brooks pointer. 885 fast_oop = new AddPNode(macro->top(), fast_oop, igvn.MakeConX(ShenandoahBrooksPointer::byte_size())); 886 macro->transform_later(fast_oop); 887 888 // Initialize Shenandoah brooks pointer to point to the object itself. 889 fast_oop_rawmem = macro->make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, ShenandoahBrooksPointer::byte_offset(), fast_oop, T_OBJECT); 890 891 return fast_oop; 892 } 893 894 // Support for GC barriers emitted during parsing 895 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const { 896 if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) { 897 return false; 898 } 899 CallLeafNode *call = node->as_CallLeaf(); 900 if (call->_name == NULL) { 901 return false; 902 } 903 904 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 || 905 strcmp(call->_name, "shenandoah_cas_obj") == 0 || 906 strcmp(call->_name, "shenandoah_wb_pre") == 0; 907 } 908 909 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const { 910 return ShenandoahBarrierNode::skip_through_barrier(c); 911 } 912 913 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, ArrayCopyPhase phase) const { 914 bool is_oop = type == T_OBJECT || type == T_ARRAY; 915 if (!is_oop) { 916 return false; 917 } 918 919 if (tightly_coupled_alloc) { 920 if (phase == Optimization) { 921 return false; 922 } 923 return !is_clone; 924 } 925 if (phase == Optimization) { 926 return !ShenandoahStoreValEnqueueBarrier; 927 } 928 return true; 929 } 930 931 bool ShenandoahBarrierSetC2::clone_needs_postbarrier(ArrayCopyNode *ac, PhaseIterGVN& igvn) { 932 Node* src = ac->in(ArrayCopyNode::Src); 933 const TypeOopPtr* src_type = igvn.type(src)->is_oopptr(); 934 if (src_type->isa_instptr() != NULL) { 935 ciInstanceKlass* ik = src_type->klass()->as_instance_klass(); 936 if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) { 937 if (ik->has_object_fields()) { 938 return true; 939 } else { 940 if (!src_type->klass_is_exact()) { 941 igvn.C->dependencies()->assert_leaf_type(ik); 942 } 943 } 944 } else { 945 return true; 946 } 947 } else if (src_type->isa_aryptr()) { 948 BasicType src_elem = src_type->klass()->as_array_klass()->element_type()->basic_type(); 949 if (src_elem == T_OBJECT || src_elem == T_ARRAY) { 950 return true; 951 } 952 } else { 953 return true; 954 } 955 return false; 956 } 957 958 void ShenandoahBarrierSetC2::clone_barrier_at_expansion(ArrayCopyNode* ac, Node* call, PhaseIterGVN& igvn) const { 959 assert(ac->is_clonebasic(), "no other kind of arraycopy here"); 960 961 if (!clone_needs_postbarrier(ac, igvn)) { 962 BarrierSetC2::clone_barrier_at_expansion(ac, call, igvn); 963 return; 964 } 965 966 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 967 Node* c = new ProjNode(call,TypeFunc::Control); 968 c = igvn.transform(c); 969 Node* m = new ProjNode(call, TypeFunc::Memory); 970 c = igvn.transform(m); 971 972 Node* dest = ac->in(ArrayCopyNode::Dest); 973 assert(dest->is_AddP(), "bad input"); 974 Node* barrier_call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(), 975 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier), 976 "shenandoah_clone_barrier", raw_adr_type); 977 barrier_call->init_req(TypeFunc::Control, c); 978 barrier_call->init_req(TypeFunc::I_O , igvn.C->top()); 979 barrier_call->init_req(TypeFunc::Memory , m); 980 barrier_call->init_req(TypeFunc::ReturnAdr, igvn.C->top()); 981 barrier_call->init_req(TypeFunc::FramePtr, igvn.C->top()); 982 barrier_call->init_req(TypeFunc::Parms+0, dest->in(AddPNode::Base)); 983 984 barrier_call = igvn.transform(barrier_call); 985 c = new ProjNode(barrier_call,TypeFunc::Control); 986 c = igvn.transform(c); 987 m = new ProjNode(barrier_call, TypeFunc::Memory); 988 m = igvn.transform(m); 989 990 Node* out_c = ac->proj_out(TypeFunc::Control); 991 Node* out_m = ac->proj_out(TypeFunc::Memory); 992 igvn.replace_node(out_c, c); 993 igvn.replace_node(out_m, m); 994 } 995 996 997 // Support for macro expanded GC barriers 998 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const { 999 if (node->Opcode() == Op_ShenandoahWriteBarrier) { 1000 state()->add_shenandoah_barrier((ShenandoahWriteBarrierNode*) node); 1001 } 1002 } 1003 1004 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const { 1005 if (node->Opcode() == Op_ShenandoahWriteBarrier) { 1006 state()->remove_shenandoah_barrier((ShenandoahWriteBarrierNode*) node); 1007 } 1008 } 1009 1010 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const { 1011 if (is_shenandoah_wb_pre_call(n)) { 1012 shenandoah_eliminate_wb_pre(n, ¯o->igvn()); 1013 } 1014 } 1015 1016 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const { 1017 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), ""); 1018 Node* c = call->as_Call()->proj_out(TypeFunc::Control); 1019 c = c->unique_ctrl_out(); 1020 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 1021 c = c->unique_ctrl_out(); 1022 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 1023 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0); 1024 assert(iff->is_If(), "expect test"); 1025 if (!is_shenandoah_marking_if(igvn, iff)) { 1026 c = c->unique_ctrl_out(); 1027 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 1028 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0); 1029 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test"); 1030 } 1031 Node* cmpx = iff->in(1)->in(1); 1032 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ)); 1033 igvn->rehash_node_delayed(call); 1034 call->del_req(call->req()-1); 1035 } 1036 1037 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const { 1038 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) { 1039 igvn->add_users_to_worklist(node); 1040 } 1041 } 1042 1043 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const { 1044 for (uint i = 0; i < useful.size(); i++) { 1045 Node* n = useful.at(i); 1046 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) { 1047 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1048 C->record_for_igvn(n->fast_out(i)); 1049 } 1050 } 1051 } 1052 for (int i = state()->shenandoah_barriers_count()-1; i >= 0; i--) { 1053 ShenandoahWriteBarrierNode* n = state()->shenandoah_barrier(i); 1054 if (!useful.member(n)) { 1055 state()->remove_shenandoah_barrier(n); 1056 } 1057 } 1058 1059 } 1060 1061 bool ShenandoahBarrierSetC2::has_special_unique_user(const Node* node) const { 1062 assert(node->outcnt() == 1, "match only for unique out"); 1063 Node* n = node->unique_out(); 1064 return node->Opcode() == Op_ShenandoahWriteBarrier && n->Opcode() == Op_ShenandoahWBMemProj; 1065 } 1066 1067 void ShenandoahBarrierSetC2::add_users_to_worklist(Unique_Node_List* worklist) const {} 1068 1069 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const { 1070 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena); 1071 } 1072 1073 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const { 1074 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state()); 1075 } 1076 1077 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be 1078 // expanded later, then now is the time to do so. 1079 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; } 1080 1081 #ifdef ASSERT 1082 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const { 1083 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeExpand) { 1084 ShenandoahBarrierNode::verify(Compile::current()->root()); 1085 } else if (phase == BarrierSetC2::BeforeCodeGen) { 1086 // Verify G1 pre-barriers 1087 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()); 1088 1089 ResourceArea *area = Thread::current()->resource_area(); 1090 Unique_Node_List visited(area); 1091 Node_List worklist(area); 1092 // We're going to walk control flow backwards starting from the Root 1093 worklist.push(compile->root()); 1094 while (worklist.size() > 0) { 1095 Node *x = worklist.pop(); 1096 if (x == NULL || x == compile->top()) continue; 1097 if (visited.member(x)) { 1098 continue; 1099 } else { 1100 visited.push(x); 1101 } 1102 1103 if (x->is_Region()) { 1104 for (uint i = 1; i < x->req(); i++) { 1105 worklist.push(x->in(i)); 1106 } 1107 } else { 1108 worklist.push(x->in(0)); 1109 // We are looking for the pattern: 1110 // /->ThreadLocal 1111 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 1112 // \->ConI(0) 1113 // We want to verify that the If and the LoadB have the same control 1114 // See GraphKit::g1_write_barrier_pre() 1115 if (x->is_If()) { 1116 IfNode *iff = x->as_If(); 1117 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 1118 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 1119 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 1120 && cmp->in(1)->is_Load()) { 1121 LoadNode *load = cmp->in(1)->as_Load(); 1122 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 1123 && load->in(2)->in(3)->is_Con() 1124 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 1125 1126 Node *if_ctrl = iff->in(0); 1127 Node *load_ctrl = load->in(0); 1128 1129 if (if_ctrl != load_ctrl) { 1130 // Skip possible CProj->NeverBranch in infinite loops 1131 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 1132 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 1133 if_ctrl = if_ctrl->in(0)->in(0); 1134 } 1135 } 1136 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 1137 } 1138 } 1139 } 1140 } 1141 } 1142 } 1143 } 1144 } 1145 #endif 1146 1147 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const { 1148 if (is_shenandoah_wb_pre_call(n)) { 1149 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt(); 1150 if (n->req() > cnt) { 1151 Node* addp = n->in(cnt); 1152 if (has_only_shenandoah_wb_pre_uses(addp)) { 1153 n->del_req(cnt); 1154 if (can_reshape) { 1155 phase->is_IterGVN()->_worklist.push(addp); 1156 } 1157 return n; 1158 } 1159 } 1160 } 1161 if (n->Opcode() == Op_CmpP) { 1162 Node* in1 = n->in(1); 1163 Node* in2 = n->in(2); 1164 if (in1->bottom_type() == TypePtr::NULL_PTR) { 1165 in2 = step_over_gc_barrier(in2); 1166 } 1167 if (in2->bottom_type() == TypePtr::NULL_PTR) { 1168 in1 = step_over_gc_barrier(in1); 1169 } 1170 PhaseIterGVN* igvn = phase->is_IterGVN(); 1171 if (in1 != n->in(1)) { 1172 if (igvn != NULL) { 1173 n->set_req_X(1, in1, igvn); 1174 } else { 1175 n->set_req(1, in1); 1176 } 1177 assert(in2 == n->in(2), "only one change"); 1178 return n; 1179 } 1180 if (in2 != n->in(2)) { 1181 if (igvn != NULL) { 1182 n->set_req_X(2, in2, igvn); 1183 } else { 1184 n->set_req(2, in2); 1185 } 1186 return n; 1187 } 1188 } else if (can_reshape && 1189 n->Opcode() == Op_If && 1190 ShenandoahWriteBarrierNode::is_heap_stable_test(n) && 1191 n->in(0) != NULL) { 1192 Node* dom = n->in(0); 1193 Node* prev_dom = n; 1194 int op = n->Opcode(); 1195 int dist = 16; 1196 // Search up the dominator tree for another heap stable test 1197 while (dom->Opcode() != op || // Not same opcode? 1198 !ShenandoahWriteBarrierNode::is_heap_stable_test(dom) || // Not same input 1? 1199 prev_dom->in(0) != dom) { // One path of test does not dominate? 1200 if (dist < 0) return NULL; 1201 1202 dist--; 1203 prev_dom = dom; 1204 dom = IfNode::up_one_dom(dom); 1205 if (!dom) return NULL; 1206 } 1207 1208 // Check that we did not follow a loop back to ourselves 1209 if (n == dom) { 1210 return NULL; 1211 } 1212 1213 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN()); 1214 } 1215 1216 return NULL; 1217 } 1218 1219 Node* ShenandoahBarrierSetC2::identity_node(PhaseGVN* phase, Node* n) const { 1220 if (n->is_Load()) { 1221 Node *mem = n->in(MemNode::Memory); 1222 Node *value = n->as_Load()->can_see_stored_value(mem, phase); 1223 if (value) { 1224 PhaseIterGVN *igvn = phase->is_IterGVN(); 1225 if (igvn != NULL && 1226 value->is_Phi() && 1227 value->req() > 2 && 1228 value->in(1) != NULL && 1229 value->in(1)->is_ShenandoahBarrier()) { 1230 if (igvn->_worklist.member(value) || 1231 igvn->_worklist.member(value->in(0)) || 1232 (value->in(0)->in(1) != NULL && 1233 value->in(0)->in(1)->is_IfProj() && 1234 (igvn->_worklist.member(value->in(0)->in(1)) || 1235 (value->in(0)->in(1)->in(0) != NULL && 1236 igvn->_worklist.member(value->in(0)->in(1)->in(0)))))) { 1237 igvn->_worklist.push(n); 1238 return n; 1239 } 1240 } 1241 // (This works even when value is a Con, but LoadNode::Value 1242 // usually runs first, producing the singleton type of the Con.) 1243 Node *value_no_barrier = step_over_gc_barrier(value->Opcode() == Op_EncodeP ? value->in(1) : value); 1244 if (value->Opcode() == Op_EncodeP) { 1245 if (value_no_barrier != value->in(1)) { 1246 Node *encode = value->clone(); 1247 encode->set_req(1, value_no_barrier); 1248 encode = phase->transform(encode); 1249 return encode; 1250 } 1251 } else { 1252 return value_no_barrier; 1253 } 1254 } 1255 } 1256 return n; 1257 } 1258 1259 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) { 1260 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1261 Node* u = n->fast_out(i); 1262 if (!is_shenandoah_wb_pre_call(u)) { 1263 return false; 1264 } 1265 } 1266 return n->outcnt() > 0; 1267 } 1268 1269 bool ShenandoahBarrierSetC2::flatten_gc_alias_type(const TypePtr*& adr_type) const { 1270 int offset = adr_type->offset(); 1271 if (offset == ShenandoahBrooksPointer::byte_offset()) { 1272 if (adr_type->isa_aryptr()) { 1273 adr_type = TypeAryPtr::make(adr_type->ptr(), adr_type->isa_aryptr()->ary(), adr_type->isa_aryptr()->klass(), false, offset); 1274 } else if (adr_type->isa_instptr()) { 1275 adr_type = TypeInstPtr::make(adr_type->ptr(), ciEnv::current()->Object_klass(), false, NULL, offset); 1276 } 1277 return true; 1278 } else { 1279 return false; 1280 } 1281 } 1282 1283 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode) const { 1284 switch (opcode) { 1285 case Op_CallLeaf: 1286 case Op_CallLeafNoFP: { 1287 assert (n->is_Call(), ""); 1288 CallNode *call = n->as_Call(); 1289 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) { 1290 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt(); 1291 if (call->req() > cnt) { 1292 assert(call->req() == cnt + 1, "only one extra input"); 1293 Node *addp = call->in(cnt); 1294 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?"); 1295 call->del_req(cnt); 1296 } 1297 } 1298 return false; 1299 } 1300 case Op_ShenandoahCompareAndSwapP: 1301 case Op_ShenandoahCompareAndSwapN: 1302 case Op_ShenandoahWeakCompareAndSwapN: 1303 case Op_ShenandoahWeakCompareAndSwapP: 1304 case Op_ShenandoahCompareAndExchangeP: 1305 case Op_ShenandoahCompareAndExchangeN: 1306 #ifdef ASSERT 1307 if( VerifyOptoOopOffsets ) { 1308 MemNode* mem = n->as_Mem(); 1309 // Check to see if address types have grounded out somehow. 1310 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 1311 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 1312 bool oop_offset_is_sane = k->contains_field_offset(tp->offset()); 1313 assert( !tp || oop_offset_is_sane, "" ); 1314 } 1315 #endif 1316 return true; 1317 case Op_ShenandoahReadBarrier: 1318 return true; 1319 case Op_ShenandoahWriteBarrier: 1320 assert(false, "should have been expanded already"); 1321 return true; 1322 default: 1323 return false; 1324 } 1325 } 1326 1327 #ifdef ASSERT 1328 bool ShenandoahBarrierSetC2::verify_gc_alias_type(const TypePtr* adr_type, int offset) const { 1329 if (offset == ShenandoahBrooksPointer::byte_offset() && 1330 (adr_type->base() == Type::AryPtr || adr_type->base() == Type::OopPtr)) { 1331 return true; 1332 } else { 1333 return false; 1334 } 1335 } 1336 #endif 1337 1338 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const { 1339 switch (opcode) { 1340 case Op_ShenandoahCompareAndExchangeP: 1341 case Op_ShenandoahCompareAndExchangeN: 1342 conn_graph->add_objload_to_connection_graph(n, delayed_worklist); 1343 // fallthrough 1344 case Op_ShenandoahWeakCompareAndSwapP: 1345 case Op_ShenandoahWeakCompareAndSwapN: 1346 case Op_ShenandoahCompareAndSwapP: 1347 case Op_ShenandoahCompareAndSwapN: 1348 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist); 1349 return true; 1350 case Op_StoreP: { 1351 Node* adr = n->in(MemNode::Address); 1352 const Type* adr_type = gvn->type(adr); 1353 // Pointer stores in G1 barriers looks like unsafe access. 1354 // Ignore such stores to be able scalar replace non-escaping 1355 // allocations. 1356 if (adr_type->isa_rawptr() && adr->is_AddP()) { 1357 Node* base = conn_graph->get_addp_base(adr); 1358 if (base->Opcode() == Op_LoadP && 1359 base->in(MemNode::Address)->is_AddP()) { 1360 adr = base->in(MemNode::Address); 1361 Node* tls = conn_graph->get_addp_base(adr); 1362 if (tls->Opcode() == Op_ThreadLocal) { 1363 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot); 1364 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()); 1365 if (offs == buf_offset) { 1366 return true; // Pre barrier previous oop value store. 1367 } 1368 } 1369 } 1370 } 1371 return false; 1372 } 1373 case Op_ShenandoahReadBarrier: 1374 case Op_ShenandoahWriteBarrier: 1375 // Barriers 'pass through' its arguments. I.e. what goes in, comes out. 1376 // It doesn't escape. 1377 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahBarrierNode::ValueIn), delayed_worklist); 1378 break; 1379 case Op_ShenandoahEnqueueBarrier: 1380 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist); 1381 break; 1382 default: 1383 // Nothing 1384 break; 1385 } 1386 return false; 1387 } 1388 1389 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const { 1390 switch (opcode) { 1391 case Op_ShenandoahCompareAndExchangeP: 1392 case Op_ShenandoahCompareAndExchangeN: { 1393 Node *adr = n->in(MemNode::Address); 1394 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL); 1395 // fallthrough 1396 } 1397 case Op_ShenandoahCompareAndSwapP: 1398 case Op_ShenandoahCompareAndSwapN: 1399 case Op_ShenandoahWeakCompareAndSwapP: 1400 case Op_ShenandoahWeakCompareAndSwapN: 1401 return conn_graph->add_final_edges_unsafe_access(n, opcode); 1402 case Op_ShenandoahReadBarrier: 1403 case Op_ShenandoahWriteBarrier: 1404 // Barriers 'pass through' its arguments. I.e. what goes in, comes out. 1405 // It doesn't escape. 1406 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahBarrierNode::ValueIn), NULL); 1407 return true; 1408 case Op_ShenandoahEnqueueBarrier: 1409 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), NULL); 1410 return true; 1411 default: 1412 // Nothing 1413 break; 1414 } 1415 return false; 1416 } 1417 1418 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const { 1419 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) || 1420 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN); 1421 1422 } 1423 1424 bool ShenandoahBarrierSetC2::escape_is_barrier_node(Node* n) const { 1425 return n->is_ShenandoahBarrier(); 1426 } 1427 1428 bool ShenandoahBarrierSetC2::matcher_find_shared_visit(Matcher* matcher, Matcher::MStack& mstack, Node* n, uint opcode, bool& mem_op, int& mem_addr_idx) const { 1429 switch (opcode) { 1430 case Op_ShenandoahReadBarrier: 1431 if (n->in(ShenandoahBarrierNode::ValueIn)->is_DecodeNarrowPtr()) { 1432 matcher->set_shared(n->in(ShenandoahBarrierNode::ValueIn)->in(1)); 1433 } 1434 matcher->set_shared(n); 1435 return true; 1436 default: 1437 break; 1438 } 1439 return false; 1440 } 1441 1442 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const { 1443 switch (opcode) { 1444 case Op_ShenandoahCompareAndExchangeP: 1445 case Op_ShenandoahCompareAndExchangeN: 1446 case Op_ShenandoahWeakCompareAndSwapP: 1447 case Op_ShenandoahWeakCompareAndSwapN: 1448 case Op_ShenandoahCompareAndSwapP: 1449 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree 1450 Node* newval = n->in(MemNode::ValueIn); 1451 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn); 1452 Node* pair = new BinaryNode(oldval, newval); 1453 n->set_req(MemNode::ValueIn,pair); 1454 n->del_req(LoadStoreConditionalNode::ExpectedIn); 1455 return true; 1456 } 1457 default: 1458 break; 1459 } 1460 return false; 1461 } 1462 1463 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const { 1464 return xop == Op_ShenandoahCompareAndExchangeP || 1465 xop == Op_ShenandoahCompareAndExchangeN || 1466 xop == Op_ShenandoahWeakCompareAndSwapP || 1467 xop == Op_ShenandoahWeakCompareAndSwapN || 1468 xop == Op_ShenandoahCompareAndSwapN || 1469 xop == Op_ShenandoahCompareAndSwapP; 1470 } 1471 1472 void ShenandoahBarrierSetC2::igvn_add_users_to_worklist(PhaseIterGVN* igvn, Node* use) const { 1473 if (use->is_ShenandoahBarrier()) { 1474 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1475 Node* u = use->fast_out(i2); 1476 Node* cmp = use->find_out_with(Op_CmpP); 1477 if (u->Opcode() == Op_CmpP) { 1478 igvn->_worklist.push(cmp); 1479 } 1480 } 1481 } 1482 } 1483 1484 void ShenandoahBarrierSetC2::ccp_analyze(PhaseCCP* ccp, Unique_Node_List& worklist, Node* use) const { 1485 if (use->is_ShenandoahBarrier()) { 1486 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1487 Node* p = use->fast_out(i2); 1488 if (p->Opcode() == Op_AddP) { 1489 for (DUIterator_Fast i3max, i3 = p->fast_outs(i3max); i3 < i3max; i3++) { 1490 Node* q = p->fast_out(i3); 1491 if (q->is_Load()) { 1492 if(q->bottom_type() != ccp->type(q)) { 1493 worklist.push(q); 1494 } 1495 } 1496 } 1497 } 1498 } 1499 } 1500 }