1 /* 2 * Copyright (c) 2015, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 */ 23 24 #include "precompiled.hpp" 25 #include "opto/compile.hpp" 26 #include "opto/castnode.hpp" 27 #include "opto/graphKit.hpp" 28 #include "opto/idealKit.hpp" 29 #include "opto/loopnode.hpp" 30 #include "opto/macro.hpp" 31 #include "opto/node.hpp" 32 #include "opto/type.hpp" 33 #include "utilities/macros.hpp" 34 #include "gc/z/c2/zBarrierSetC2.hpp" 35 #include "gc/z/zThreadLocalData.hpp" 36 #include "gc/z/zBarrierSetRuntime.hpp" 37 38 ZBarrierSetC2State::ZBarrierSetC2State(Arena* comp_arena) 39 : _load_barrier_nodes(new (comp_arena) GrowableArray<LoadBarrierNode*>(comp_arena, 8, 0, NULL)) {} 40 41 int ZBarrierSetC2State::load_barrier_count() const { 42 return _load_barrier_nodes->length(); 43 } 44 45 void ZBarrierSetC2State::add_load_barrier_node(LoadBarrierNode * n) { 46 assert(!_load_barrier_nodes->contains(n), " duplicate entry in expand list"); 47 _load_barrier_nodes->append(n); 48 } 49 50 void ZBarrierSetC2State::remove_load_barrier_node(LoadBarrierNode * n) { 51 // this function may be called twice for a node so check 52 // that the node is in the array before attempting to remove it 53 if (_load_barrier_nodes->contains(n)) { 54 _load_barrier_nodes->remove(n); 55 } 56 } 57 58 LoadBarrierNode* ZBarrierSetC2State::load_barrier_node(int idx) const { 59 return _load_barrier_nodes->at(idx); 60 } 61 62 void* ZBarrierSetC2::create_barrier_state(Arena* comp_arena) const { 63 return new(comp_arena) ZBarrierSetC2State(comp_arena); 64 } 65 66 ZBarrierSetC2State* ZBarrierSetC2::state() const { 67 return reinterpret_cast<ZBarrierSetC2State*>(Compile::current()->barrier_set_state()); 68 } 69 70 bool ZBarrierSetC2::is_gc_barrier_node(Node* node) const { 71 // 1. This step follows potential oop projections of a load barrier before expansion 72 if (node->is_Proj()) { 73 node = node->in(0); 74 } 75 76 // 2. This step checks for unexpanded load barriers 77 if (node->is_LoadBarrier()) { 78 return true; 79 } 80 81 // 3. This step checks for the phi corresponding to an optimized load barrier expansion 82 if (node->is_Phi()) { 83 PhiNode* phi = node->as_Phi(); 84 Node* n = phi->in(1); 85 if (n != NULL && (n->is_LoadBarrierSlowReg() || n->is_LoadBarrierWeakSlowReg())) { 86 return true; 87 } 88 } 89 90 return false; 91 } 92 93 void ZBarrierSetC2::register_potential_barrier_node(Node* node) const { 94 if (node->is_LoadBarrier()) { 95 state()->add_load_barrier_node(node->as_LoadBarrier()); 96 } 97 } 98 99 void ZBarrierSetC2::unregister_potential_barrier_node(Node* node) const { 100 if (node->is_LoadBarrier()) { 101 state()->remove_load_barrier_node(node->as_LoadBarrier()); 102 } 103 } 104 105 void ZBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const { 106 // Remove useless LoadBarrier nodes 107 ZBarrierSetC2State* s = state(); 108 for (int i = s->load_barrier_count()-1; i >= 0; i--) { 109 LoadBarrierNode* n = s->load_barrier_node(i); 110 if (!useful.member(n)) { 111 unregister_potential_barrier_node(n); 112 } 113 } 114 } 115 116 void ZBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const { 117 if (node->is_LoadBarrier() && !node->as_LoadBarrier()->has_true_uses()) { 118 igvn->_worklist.push(node); 119 } 120 } 121 122 void ZBarrierSetC2::find_dominating_barriers(PhaseIterGVN& igvn) { 123 // Look for dominating barriers on the same address only once all 124 // other loop opts are over: loop opts may cause a safepoint to be 125 // inserted between a barrier and its dominating barrier. 126 Compile* C = Compile::current(); 127 ZBarrierSetC2* bs = (ZBarrierSetC2*)BarrierSet::barrier_set()->barrier_set_c2(); 128 ZBarrierSetC2State* s = bs->state(); 129 if (s->load_barrier_count() >= 2) { 130 Compile::TracePhase tp("idealLoop", &C->timers[Phase::_t_idealLoop]); 131 PhaseIdealLoop ideal_loop(igvn, LoopOptsLastRound); 132 if (C->major_progress()) C->print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 133 } 134 } 135 136 void ZBarrierSetC2::add_users_to_worklist(Unique_Node_List* worklist) const { 137 // Permanent temporary workaround 138 // Loadbarriers may have non-obvious dead uses keeping them alive during parsing. The use is 139 // removed by RemoveUseless (after parsing, before optimize) but the barriers won't be added to 140 // the worklist. Unless we add them explicitly they are not guaranteed to end up there. 141 ZBarrierSetC2State* s = state(); 142 143 for (int i = 0; i < s->load_barrier_count(); i++) { 144 LoadBarrierNode* n = s->load_barrier_node(i); 145 worklist->push(n); 146 } 147 } 148 149 const TypeFunc* ZBarrierSetC2::load_barrier_Type() const { 150 const Type** fields; 151 152 // Create input types (domain) 153 fields = TypeTuple::fields(2); 154 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; 155 fields[TypeFunc::Parms+1] = TypeOopPtr::BOTTOM; 156 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields); 157 158 // Create result type (range) 159 fields = TypeTuple::fields(1); 160 fields[TypeFunc::Parms+0] = TypeInstPtr::BOTTOM; 161 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields); 162 163 return TypeFunc::make(domain, range); 164 } 165 166 // == LoadBarrierNode == 167 168 LoadBarrierNode::LoadBarrierNode(Compile* C, 169 Node* c, 170 Node* mem, 171 Node* val, 172 Node* adr, 173 bool weak, 174 bool writeback, 175 bool oop_reload_allowed) : 176 MultiNode(Number_of_Inputs), 177 _weak(weak), 178 _writeback(writeback), 179 _oop_reload_allowed(oop_reload_allowed) { 180 init_req(Control, c); 181 init_req(Memory, mem); 182 init_req(Oop, val); 183 init_req(Address, adr); 184 init_req(Similar, C->top()); 185 186 init_class_id(Class_LoadBarrier); 187 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 188 bs->register_potential_barrier_node(this); 189 } 190 191 const Type *LoadBarrierNode::bottom_type() const { 192 const Type** floadbarrier = (const Type **)(Compile::current()->type_arena()->Amalloc_4((Number_of_Outputs)*sizeof(Type*))); 193 Node* in_oop = in(Oop); 194 floadbarrier[Control] = Type::CONTROL; 195 floadbarrier[Memory] = Type::MEMORY; 196 floadbarrier[Oop] = in_oop == NULL ? Type::TOP : in_oop->bottom_type(); 197 return TypeTuple::make(Number_of_Outputs, floadbarrier); 198 } 199 200 const Type *LoadBarrierNode::Value(PhaseGVN *phase) const { 201 const Type** floadbarrier = (const Type **)(phase->C->type_arena()->Amalloc_4((Number_of_Outputs)*sizeof(Type*))); 202 const Type* val_t = phase->type(in(Oop)); 203 floadbarrier[Control] = Type::CONTROL; 204 floadbarrier[Memory] = Type::MEMORY; 205 floadbarrier[Oop] = val_t; 206 return TypeTuple::make(Number_of_Outputs, floadbarrier); 207 } 208 209 bool LoadBarrierNode::is_dominator(PhaseIdealLoop* phase, bool linear_only, Node *d, Node *n) { 210 if (phase != NULL) { 211 return phase->is_dominator(d, n); 212 } 213 214 for (int i = 0; i < 10 && n != NULL; i++) { 215 n = IfNode::up_one_dom(n, linear_only); 216 if (n == d) { 217 return true; 218 } 219 } 220 221 return false; 222 } 223 224 LoadBarrierNode* LoadBarrierNode::has_dominating_barrier(PhaseIdealLoop* phase, bool linear_only, bool look_for_similar) { 225 Node* val = in(LoadBarrierNode::Oop); 226 if (in(Similar)->is_Proj() && in(Similar)->in(0)->is_LoadBarrier()) { 227 LoadBarrierNode* lb = in(Similar)->in(0)->as_LoadBarrier(); 228 assert(lb->in(Address) == in(Address), ""); 229 // Load barrier on Similar edge dominates so if it now has the Oop field it can replace this barrier. 230 if (lb->in(Oop) == in(Oop)) { 231 return lb; 232 } 233 // Follow chain of load barrier through Similar edges 234 while (!lb->in(Similar)->is_top()) { 235 lb = lb->in(Similar)->in(0)->as_LoadBarrier(); 236 assert(lb->in(Address) == in(Address), ""); 237 } 238 if (lb != in(Similar)->in(0)) { 239 return lb; 240 } 241 } 242 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { 243 Node* u = val->fast_out(i); 244 if (u != this && u->is_LoadBarrier() && u->in(Oop) == val && u->as_LoadBarrier()->has_true_uses()) { 245 Node* this_ctrl = in(LoadBarrierNode::Control); 246 Node* other_ctrl = u->in(LoadBarrierNode::Control); 247 if (is_dominator(phase, linear_only, other_ctrl, this_ctrl)) { 248 return u->as_LoadBarrier(); 249 } 250 } 251 } 252 253 if (ZVerifyLoadBarriers || can_be_eliminated()) { 254 return NULL; 255 } 256 257 if (!look_for_similar) { 258 return NULL; 259 } 260 261 Node* addr = in(LoadBarrierNode::Address); 262 for (DUIterator_Fast imax, i = addr->fast_outs(imax); i < imax; i++) { 263 Node* u = addr->fast_out(i); 264 if (u != this && u->is_LoadBarrier() && u->as_LoadBarrier()->has_true_uses()) { 265 Node* this_ctrl = in(LoadBarrierNode::Control); 266 Node* other_ctrl = u->in(LoadBarrierNode::Control); 267 if (is_dominator(phase, linear_only, other_ctrl, this_ctrl)) { 268 ResourceMark rm; 269 Unique_Node_List wq; 270 wq.push(in(LoadBarrierNode::Control)); 271 bool ok = true; 272 bool dom_found = false; 273 for (uint next = 0; next < wq.size(); ++next) { 274 Node *n = wq.at(next); 275 if (n->is_top()) { 276 return NULL; 277 } 278 assert(n->is_CFG(), ""); 279 if (n->is_SafePoint()) { 280 ok = false; 281 break; 282 } 283 if (n == u) { 284 dom_found = true; 285 continue; 286 } 287 if (n->is_Region()) { 288 for (uint i = 1; i < n->req(); i++) { 289 Node* m = n->in(i); 290 if (m != NULL) { 291 wq.push(m); 292 } 293 } 294 } else { 295 Node* m = n->in(0); 296 if (m != NULL) { 297 wq.push(m); 298 } 299 } 300 } 301 if (ok) { 302 assert(dom_found, ""); 303 return u->as_LoadBarrier();; 304 } 305 break; 306 } 307 } 308 } 309 310 return NULL; 311 } 312 313 void LoadBarrierNode::push_dominated_barriers(PhaseIterGVN* igvn) const { 314 // Change to that barrier may affect a dominated barrier so re-push those 315 Node* val = in(LoadBarrierNode::Oop); 316 317 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { 318 Node* u = val->fast_out(i); 319 if (u != this && u->is_LoadBarrier() && u->in(Oop) == val) { 320 Node* this_ctrl = in(Control); 321 Node* other_ctrl = u->in(Control); 322 if (is_dominator(NULL, false, this_ctrl, other_ctrl)) { 323 igvn->_worklist.push(u); 324 } 325 } 326 327 Node* addr = in(LoadBarrierNode::Address); 328 for (DUIterator_Fast imax, i = addr->fast_outs(imax); i < imax; i++) { 329 Node* u = addr->fast_out(i); 330 if (u != this && u->is_LoadBarrier() && u->in(Similar)->is_top()) { 331 Node* this_ctrl = in(Control); 332 Node* other_ctrl = u->in(Control); 333 if (is_dominator(NULL, false, this_ctrl, other_ctrl)) { 334 igvn->_worklist.push(u); 335 } 336 } 337 } 338 } 339 } 340 341 Node *LoadBarrierNode::Identity(PhaseGVN *phase) { 342 if (!phase->C->directive()->ZOptimizeLoadBarriersOption) { 343 return this; 344 } 345 346 bool redundant_addr = false; 347 LoadBarrierNode* dominating_barrier = has_dominating_barrier(NULL, true, false); 348 if (dominating_barrier != NULL) { 349 assert(dominating_barrier->in(Oop) == in(Oop), ""); 350 return dominating_barrier; 351 } 352 353 return this; 354 } 355 356 Node *LoadBarrierNode::Ideal(PhaseGVN *phase, bool can_reshape) { 357 if (remove_dead_region(phase, can_reshape)) { 358 return this; 359 } 360 361 Node* val = in(Oop); 362 Node* mem = in(Memory); 363 Node* ctrl = in(Control); 364 Node* adr = in(Address); 365 assert(val->Opcode() != Op_LoadN, ""); 366 367 if (mem->is_MergeMem()) { 368 Node* new_mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 369 set_req(Memory, new_mem); 370 if (mem->outcnt() == 0 && can_reshape) { 371 phase->is_IterGVN()->_worklist.push(mem); 372 } 373 374 return this; 375 } 376 377 bool optimizeLoadBarriers = phase->C->directive()->ZOptimizeLoadBarriersOption; 378 LoadBarrierNode* dominating_barrier = optimizeLoadBarriers ? has_dominating_barrier(NULL, !can_reshape, !phase->C->major_progress()) : NULL; 379 if (dominating_barrier != NULL && dominating_barrier->in(Oop) != in(Oop)) { 380 assert(in(Address) == dominating_barrier->in(Address), ""); 381 set_req(Similar, dominating_barrier->proj_out(Oop)); 382 return this; 383 } 384 385 bool eliminate = (optimizeLoadBarriers && !(val->is_Phi() || val->Opcode() == Op_LoadP || val->Opcode() == Op_GetAndSetP || val->is_DecodeN())) || 386 (can_reshape && (dominating_barrier != NULL || !has_true_uses())); 387 388 if (eliminate) { 389 if (can_reshape) { 390 PhaseIterGVN* igvn = phase->is_IterGVN(); 391 Node* out_ctrl = proj_out_or_null(Control); 392 Node* out_res = proj_out_or_null(Oop); 393 394 if (out_ctrl != NULL) { 395 igvn->replace_node(out_ctrl, ctrl); 396 } 397 398 // That transformation may cause the Similar edge on the load barrier to be invalid 399 fix_similar_in_uses(igvn); 400 if (out_res != NULL) { 401 if (dominating_barrier != NULL) { 402 igvn->replace_node(out_res, dominating_barrier->proj_out(Oop)); 403 } else { 404 igvn->replace_node(out_res, val); 405 } 406 } 407 } 408 409 return new ConINode(TypeInt::ZERO); 410 } 411 412 // If the Similar edge is no longer a load barrier, clear it 413 Node* similar = in(Similar); 414 if (!similar->is_top() && !(similar->is_Proj() && similar->in(0)->is_LoadBarrier())) { 415 set_req(Similar, phase->C->top()); 416 return this; 417 } 418 419 if (can_reshape) { 420 // If this barrier is linked through the Similar edge by a 421 // dominated barrier and both barriers have the same Oop field, 422 // the dominated barrier can go away, so push it for reprocessing. 423 // We also want to avoid a barrier to depend on another dominating 424 // barrier through its Similar edge that itself depend on another 425 // barrier through its Similar edge and rather have the first 426 // depend on the third. 427 PhaseIterGVN* igvn = phase->is_IterGVN(); 428 Node* out_res = proj_out(Oop); 429 for (DUIterator_Fast imax, i = out_res->fast_outs(imax); i < imax; i++) { 430 Node* u = out_res->fast_out(i); 431 if (u->is_LoadBarrier() && u->in(Similar) == out_res && 432 (u->in(Oop) == val || !u->in(Similar)->is_top())) { 433 igvn->_worklist.push(u); 434 } 435 } 436 437 push_dominated_barriers(igvn); 438 } 439 440 return NULL; 441 } 442 443 void LoadBarrierNode::fix_similar_in_uses(PhaseIterGVN* igvn) { 444 Node* out_res = proj_out_or_null(Oop); 445 if (out_res == NULL) { 446 return; 447 } 448 449 for (DUIterator_Fast imax, i = out_res->fast_outs(imax); i < imax; i++) { 450 Node* u = out_res->fast_out(i); 451 if (u->is_LoadBarrier() && u->in(Similar) == out_res) { 452 igvn->replace_input_of(u, Similar, igvn->C->top()); 453 --i; 454 --imax; 455 } 456 } 457 } 458 459 bool LoadBarrierNode::has_true_uses() const { 460 Node* out_res = proj_out_or_null(Oop); 461 if (out_res == NULL) { 462 return false; 463 } 464 465 for (DUIterator_Fast imax, i = out_res->fast_outs(imax); i < imax; i++) { 466 Node* u = out_res->fast_out(i); 467 if (!u->is_LoadBarrier() || u->in(Similar) != out_res) { 468 return true; 469 } 470 } 471 472 return false; 473 } 474 475 // == Accesses == 476 477 Node* ZBarrierSetC2::make_cas_loadbarrier(C2AtomicParseAccess& access) const { 478 assert(!UseCompressedOops, "Not allowed"); 479 CompareAndSwapNode* cas = (CompareAndSwapNode*)access.raw_access(); 480 PhaseGVN& gvn = access.gvn(); 481 Compile* C = Compile::current(); 482 GraphKit* kit = access.kit(); 483 484 Node* in_ctrl = cas->in(MemNode::Control); 485 Node* in_mem = cas->in(MemNode::Memory); 486 Node* in_adr = cas->in(MemNode::Address); 487 Node* in_val = cas->in(MemNode::ValueIn); 488 Node* in_expected = cas->in(LoadStoreConditionalNode::ExpectedIn); 489 490 float likely = PROB_LIKELY(0.999); 491 492 const TypePtr *adr_type = gvn.type(in_adr)->isa_ptr(); 493 Compile::AliasType* alias_type = C->alias_type(adr_type); 494 int alias_idx = C->get_alias_index(adr_type); 495 496 // Outer check - true: continue, false: load and check 497 Node* region = new RegionNode(3); 498 Node* phi = new PhiNode(region, TypeInt::BOOL); 499 Node* phi_mem = new PhiNode(region, Type::MEMORY, adr_type); 500 501 // Inner check - is the healed ref equal to the expected 502 Node* region2 = new RegionNode(3); 503 Node* phi2 = new PhiNode(region2, TypeInt::BOOL); 504 Node* phi_mem2 = new PhiNode(region2, Type::MEMORY, adr_type); 505 506 // CAS node returns 0 or 1 507 Node* cmp = gvn.transform(new CmpINode(cas, kit->intcon(0))); 508 Node* bol = gvn.transform(new BoolNode(cmp, BoolTest::ne))->as_Bool(); 509 IfNode* iff = gvn.transform(new IfNode(in_ctrl, bol, likely, COUNT_UNKNOWN))->as_If(); 510 Node* then = gvn.transform(new IfTrueNode(iff)); 511 Node* elsen = gvn.transform(new IfFalseNode(iff)); 512 513 Node* scmemproj1 = gvn.transform(new SCMemProjNode(cas)); 514 515 kit->set_memory(scmemproj1, alias_idx); 516 phi_mem->init_req(1, scmemproj1); 517 phi_mem2->init_req(2, scmemproj1); 518 519 // CAS fail - reload and heal oop 520 Node* reload = kit->make_load(elsen, in_adr, TypeOopPtr::BOTTOM, T_OBJECT, MemNode::unordered); 521 Node* barrier = gvn.transform(new LoadBarrierNode(C, elsen, scmemproj1, reload, in_adr, false, true, false)); 522 Node* barrierctrl = gvn.transform(new ProjNode(barrier, LoadBarrierNode::Control)); 523 Node* barrierdata = gvn.transform(new ProjNode(barrier, LoadBarrierNode::Oop)); 524 525 // Check load 526 Node* tmpX = gvn.transform(new CastP2XNode(NULL, barrierdata)); 527 Node* in_expX = gvn.transform(new CastP2XNode(NULL, in_expected)); 528 Node* cmp2 = gvn.transform(new CmpXNode(tmpX, in_expX)); 529 Node *bol2 = gvn.transform(new BoolNode(cmp2, BoolTest::ne))->as_Bool(); 530 IfNode* iff2 = gvn.transform(new IfNode(barrierctrl, bol2, likely, COUNT_UNKNOWN))->as_If(); 531 Node* then2 = gvn.transform(new IfTrueNode(iff2)); 532 Node* elsen2 = gvn.transform(new IfFalseNode(iff2)); 533 534 // redo CAS 535 Node* cas2 = gvn.transform(new CompareAndSwapPNode(elsen2, kit->memory(alias_idx), in_adr, in_val, in_expected, cas->order())); 536 Node* scmemproj2 = gvn.transform(new SCMemProjNode(cas2)); 537 kit->set_control(elsen2); 538 kit->set_memory(scmemproj2, alias_idx); 539 540 // Merge inner flow - check if healed oop was equal too expected. 541 region2->set_req(1, kit->control()); 542 region2->set_req(2, then2); 543 phi2->set_req(1, cas2); 544 phi2->set_req(2, kit->intcon(0)); 545 phi_mem2->init_req(1, scmemproj2); 546 kit->set_memory(phi_mem2, alias_idx); 547 548 // Merge outer flow - then check if first CAS succeeded 549 region->set_req(1, then); 550 region->set_req(2, region2); 551 phi->set_req(1, kit->intcon(1)); 552 phi->set_req(2, phi2); 553 phi_mem->init_req(2, phi_mem2); 554 kit->set_memory(phi_mem, alias_idx); 555 556 gvn.transform(region2); 557 gvn.transform(phi2); 558 gvn.transform(phi_mem2); 559 gvn.transform(region); 560 gvn.transform(phi); 561 gvn.transform(phi_mem); 562 563 kit->set_control(region); 564 kit->insert_mem_bar(Op_MemBarCPUOrder); 565 566 return phi; 567 } 568 569 Node* ZBarrierSetC2::make_cmpx_loadbarrier(C2AtomicParseAccess& access) const { 570 CompareAndExchangePNode* cmpx = (CompareAndExchangePNode*)access.raw_access(); 571 GraphKit* kit = access.kit(); 572 PhaseGVN& gvn = kit->gvn(); 573 Compile* C = Compile::current(); 574 575 Node* in_ctrl = cmpx->in(MemNode::Control); 576 Node* in_mem = cmpx->in(MemNode::Memory); 577 Node* in_adr = cmpx->in(MemNode::Address); 578 Node* in_val = cmpx->in(MemNode::ValueIn); 579 Node* in_expected = cmpx->in(LoadStoreConditionalNode::ExpectedIn); 580 581 float likely = PROB_LIKELY(0.999); 582 583 const TypePtr *adr_type = cmpx->get_ptr_type(); 584 Compile::AliasType* alias_type = C->alias_type(adr_type); 585 int alias_idx = C->get_alias_index(adr_type); 586 587 // Outer check - true: continue, false: load and check 588 Node* region = new RegionNode(3); 589 Node* phi = new PhiNode(region, adr_type); 590 591 // Inner check - is the healed ref equal to the expected 592 Node* region2 = new RegionNode(3); 593 Node* phi2 = new PhiNode(region2, adr_type); 594 595 // Check if cmpx succeeded 596 Node* cmp = gvn.transform(new CmpPNode(cmpx, in_expected)); 597 Node* bol = gvn.transform(new BoolNode(cmp, BoolTest::eq))->as_Bool(); 598 IfNode* iff = gvn.transform(new IfNode(in_ctrl, bol, likely, COUNT_UNKNOWN))->as_If(); 599 Node* then = gvn.transform(new IfTrueNode(iff)); 600 Node* elsen = gvn.transform(new IfFalseNode(iff)); 601 602 Node* scmemproj1 = gvn.transform(new SCMemProjNode(cmpx)); 603 kit->set_memory(scmemproj1, alias_idx); 604 605 // CAS fail - reload and heal oop 606 Node* reload = kit->make_load(elsen, in_adr, TypeOopPtr::BOTTOM, T_OBJECT, MemNode::unordered); 607 Node* barrier = gvn.transform(new LoadBarrierNode(C, elsen, scmemproj1, reload, in_adr, false, true, false)); 608 Node* barrierctrl = gvn.transform(new ProjNode(barrier, LoadBarrierNode::Control)); 609 Node* barrierdata = gvn.transform(new ProjNode(barrier, LoadBarrierNode::Oop)); 610 611 // Check load 612 Node* tmpX = gvn.transform(new CastP2XNode(NULL, barrierdata)); 613 Node* in_expX = gvn.transform(new CastP2XNode(NULL, in_expected)); 614 Node* cmp2 = gvn.transform(new CmpXNode(tmpX, in_expX)); 615 Node *bol2 = gvn.transform(new BoolNode(cmp2, BoolTest::ne))->as_Bool(); 616 IfNode* iff2 = gvn.transform(new IfNode(barrierctrl, bol2, likely, COUNT_UNKNOWN))->as_If(); 617 Node* then2 = gvn.transform(new IfTrueNode(iff2)); 618 Node* elsen2 = gvn.transform(new IfFalseNode(iff2)); 619 620 // Redo CAS 621 Node* cmpx2 = gvn.transform(new CompareAndExchangePNode(elsen2, kit->memory(alias_idx), in_adr, in_val, in_expected, adr_type, cmpx->get_ptr_type(), cmpx->order())); 622 Node* scmemproj2 = gvn.transform(new SCMemProjNode(cmpx2)); 623 kit->set_control(elsen2); 624 kit->set_memory(scmemproj2, alias_idx); 625 626 // Merge inner flow - check if healed oop was equal too expected. 627 region2->set_req(1, kit->control()); 628 region2->set_req(2, then2); 629 phi2->set_req(1, cmpx2); 630 phi2->set_req(2, barrierdata); 631 632 // Merge outer flow - then check if first cas succeeded 633 region->set_req(1, then); 634 region->set_req(2, region2); 635 phi->set_req(1, cmpx); 636 phi->set_req(2, phi2); 637 638 gvn.transform(region2); 639 gvn.transform(phi2); 640 gvn.transform(region); 641 gvn.transform(phi); 642 643 kit->set_control(region); 644 kit->set_memory(in_mem, alias_idx); 645 kit->insert_mem_bar(Op_MemBarCPUOrder); 646 647 return phi; 648 } 649 650 Node* ZBarrierSetC2::load_barrier(GraphKit* kit, Node* val, Node* adr, bool weak, bool writeback, bool oop_reload_allowed) const { 651 PhaseGVN& gvn = kit->gvn(); 652 Node* barrier = new LoadBarrierNode(Compile::current(), kit->control(), kit->memory(TypeRawPtr::BOTTOM), val, adr, weak, writeback, oop_reload_allowed); 653 Node* transformed_barrier = gvn.transform(barrier); 654 655 if (transformed_barrier->is_LoadBarrier()) { 656 if (barrier == transformed_barrier) { 657 kit->set_control(gvn.transform(new ProjNode(barrier, LoadBarrierNode::Control))); 658 } 659 Node* result = gvn.transform(new ProjNode(transformed_barrier, LoadBarrierNode::Oop)); 660 assert(is_gc_barrier_node(result), "sanity"); 661 assert(step_over_gc_barrier(result) == val, "sanity"); 662 return result; 663 } else { 664 return val; 665 } 666 } 667 668 static bool barrier_needed(C2Access& access) { 669 return ZBarrierSet::barrier_needed(access.decorators(), access.type()); 670 } 671 672 Node* ZBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const { 673 Node* p = BarrierSetC2::load_at_resolved(access, val_type); 674 if (!barrier_needed(access)) { 675 return p; 676 } 677 678 bool weak = (access.decorators() & ON_WEAK_OOP_REF) != 0; 679 680 assert(access.is_parse_access(), "entry not supported at optimization time"); 681 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access); 682 GraphKit* kit = parse_access.kit(); 683 PhaseGVN& gvn = kit->gvn(); 684 Node* adr = access.addr().node(); 685 Node* heap_base_oop = access.base(); 686 bool unsafe = (access.decorators() & C2_UNSAFE_ACCESS) != 0; 687 if (unsafe) { 688 if (!ZVerifyLoadBarriers) { 689 p = load_barrier(kit, p, adr); 690 } else { 691 if (!TypePtr::NULL_PTR->higher_equal(gvn.type(heap_base_oop))) { 692 p = load_barrier(kit, p, adr); 693 } else { 694 IdealKit ideal(kit); 695 IdealVariable res(ideal); 696 #define __ ideal. 697 __ declarations_done(); 698 __ set(res, p); 699 __ if_then(heap_base_oop, BoolTest::ne, kit->null(), PROB_UNLIKELY(0.999)); { 700 kit->sync_kit(ideal); 701 p = load_barrier(kit, p, adr); 702 __ set(res, p); 703 __ sync_kit(kit); 704 } __ end_if(); 705 kit->final_sync(ideal); 706 p = __ value(res); 707 #undef __ 708 } 709 } 710 return p; 711 } else { 712 return load_barrier(parse_access.kit(), p, access.addr().node(), weak, true, true); 713 } 714 } 715 716 Node* ZBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val, 717 Node* new_val, const Type* val_type) const { 718 Node* result = BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, val_type); 719 if (!barrier_needed(access)) { 720 return result; 721 } 722 723 access.set_needs_pinning(false); 724 return make_cmpx_loadbarrier(access); 725 } 726 727 Node* ZBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val, 728 Node* new_val, const Type* value_type) const { 729 Node* result = BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type); 730 if (!barrier_needed(access)) { 731 return result; 732 } 733 734 Node* load_store = access.raw_access(); 735 bool weak_cas = (access.decorators() & C2_WEAK_CMPXCHG) != 0; 736 bool expected_is_null = (expected_val->get_ptr_type() == TypePtr::NULL_PTR); 737 738 if (!expected_is_null) { 739 if (weak_cas) { 740 access.set_needs_pinning(false); 741 load_store = make_cas_loadbarrier(access); 742 } else { 743 access.set_needs_pinning(false); 744 load_store = make_cas_loadbarrier(access); 745 } 746 } 747 748 return load_store; 749 } 750 751 Node* ZBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* new_val, const Type* val_type) const { 752 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, new_val, val_type); 753 if (!barrier_needed(access)) { 754 return result; 755 } 756 757 Node* load_store = access.raw_access(); 758 Node* adr = access.addr().node(); 759 760 assert(access.is_parse_access(), "entry not supported at optimization time"); 761 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access); 762 return load_barrier(parse_access.kit(), load_store, adr, false, false, false); 763 } 764 765 // == Macro Expansion == 766 767 void ZBarrierSetC2::expand_loadbarrier_node(PhaseMacroExpand* phase, LoadBarrierNode* barrier) const { 768 Node* in_ctrl = barrier->in(LoadBarrierNode::Control); 769 Node* in_mem = barrier->in(LoadBarrierNode::Memory); 770 Node* in_val = barrier->in(LoadBarrierNode::Oop); 771 Node* in_adr = barrier->in(LoadBarrierNode::Address); 772 773 Node* out_ctrl = barrier->proj_out(LoadBarrierNode::Control); 774 Node* out_res = barrier->proj_out(LoadBarrierNode::Oop); 775 776 PhaseIterGVN &igvn = phase->igvn(); 777 778 if (ZVerifyLoadBarriers) { 779 igvn.replace_node(out_res, in_val); 780 igvn.replace_node(out_ctrl, in_ctrl); 781 return; 782 } 783 784 if (barrier->can_be_eliminated()) { 785 // Clone and pin the load for this barrier below the dominating 786 // barrier: the load cannot be allowed to float above the 787 // dominating barrier 788 Node* load = in_val; 789 790 if (load->is_Load()) { 791 Node* new_load = load->clone(); 792 Node* addp = new_load->in(MemNode::Address); 793 assert(addp->is_AddP() || addp->is_Phi() || addp->is_Load(), "bad address"); 794 Node* cast = new CastPPNode(addp, igvn.type(addp), true); 795 Node* ctrl = NULL; 796 Node* similar = barrier->in(LoadBarrierNode::Similar); 797 if (similar->is_Phi()) { 798 // already expanded 799 ctrl = similar->in(0); 800 } else { 801 assert(similar->is_Proj() && similar->in(0)->is_LoadBarrier(), "unexpected graph shape"); 802 ctrl = similar->in(0)->as_LoadBarrier()->proj_out(LoadBarrierNode::Control); 803 } 804 assert(ctrl != NULL, "bad control"); 805 cast->set_req(0, ctrl); 806 igvn.transform(cast); 807 new_load->set_req(MemNode::Address, cast); 808 igvn.transform(new_load); 809 810 igvn.replace_node(out_res, new_load); 811 igvn.replace_node(out_ctrl, in_ctrl); 812 return; 813 } 814 // cannot eliminate 815 } 816 817 // There are two cases that require the basic loadbarrier 818 // 1) When the writeback of a healed oop must be avoided (swap) 819 // 2) When we must guarantee that no reload of is done (swap, cas, cmpx) 820 if (!barrier->is_writeback()) { 821 assert(!barrier->oop_reload_allowed(), "writeback barriers should be marked as requires oop"); 822 } 823 824 if (!barrier->oop_reload_allowed()) { 825 expand_loadbarrier_basic(phase, barrier); 826 } else { 827 expand_loadbarrier_optimized(phase, barrier); 828 } 829 } 830 831 // Basic loadbarrier using conventional argument passing 832 void ZBarrierSetC2::expand_loadbarrier_basic(PhaseMacroExpand* phase, LoadBarrierNode *barrier) const { 833 PhaseIterGVN &igvn = phase->igvn(); 834 835 Node* in_ctrl = barrier->in(LoadBarrierNode::Control); 836 Node* in_mem = barrier->in(LoadBarrierNode::Memory); 837 Node* in_val = barrier->in(LoadBarrierNode::Oop); 838 Node* in_adr = barrier->in(LoadBarrierNode::Address); 839 840 Node* out_ctrl = barrier->proj_out(LoadBarrierNode::Control); 841 Node* out_res = barrier->proj_out(LoadBarrierNode::Oop); 842 843 float unlikely = PROB_UNLIKELY(0.999); 844 const Type* in_val_maybe_null_t = igvn.type(in_val); 845 846 Node* jthread = igvn.transform(new ThreadLocalNode()); 847 Node* adr = phase->basic_plus_adr(jthread, in_bytes(ZThreadLocalData::address_bad_mask_offset())); 848 Node* bad_mask = igvn.transform(LoadNode::make(igvn, in_ctrl, in_mem, adr, TypeRawPtr::BOTTOM, TypeX_X, TypeX_X->basic_type(), MemNode::unordered)); 849 Node* cast = igvn.transform(new CastP2XNode(in_ctrl, in_val)); 850 Node* obj_masked = igvn.transform(new AndXNode(cast, bad_mask)); 851 Node* cmp = igvn.transform(new CmpXNode(obj_masked, igvn.zerocon(TypeX_X->basic_type()))); 852 Node *bol = igvn.transform(new BoolNode(cmp, BoolTest::ne))->as_Bool(); 853 IfNode* iff = igvn.transform(new IfNode(in_ctrl, bol, unlikely, COUNT_UNKNOWN))->as_If(); 854 Node* then = igvn.transform(new IfTrueNode(iff)); 855 Node* elsen = igvn.transform(new IfFalseNode(iff)); 856 857 Node* result_region; 858 Node* result_val; 859 860 result_region = new RegionNode(3); 861 result_val = new PhiNode(result_region, TypeInstPtr::BOTTOM); 862 863 result_region->set_req(1, elsen); 864 Node* res = igvn.transform(new CastPPNode(in_val, in_val_maybe_null_t)); 865 res->init_req(0, elsen); 866 result_val->set_req(1, res); 867 868 const TypeFunc *tf = load_barrier_Type(); 869 Node* call; 870 if (barrier->is_weak()) { 871 call = new CallLeafNode(tf, 872 ZBarrierSetRuntime::load_barrier_on_weak_oop_field_preloaded_addr(), 873 "ZBarrierSetRuntime::load_barrier_on_weak_oop_field_preloaded", 874 TypeRawPtr::BOTTOM); 875 } else { 876 call = new CallLeafNode(tf, 877 ZBarrierSetRuntime::load_barrier_on_oop_field_preloaded_addr(), 878 "ZBarrierSetRuntime::load_barrier_on_oop_field_preloaded", 879 TypeRawPtr::BOTTOM); 880 } 881 882 call->init_req(TypeFunc::Control, then); 883 call->init_req(TypeFunc::I_O , phase->top()); 884 call->init_req(TypeFunc::Memory , in_mem); 885 call->init_req(TypeFunc::FramePtr, phase->top()); 886 call->init_req(TypeFunc::ReturnAdr, phase->top()); 887 call->init_req(TypeFunc::Parms+0, in_val); 888 if (barrier->is_writeback()) { 889 call->init_req(TypeFunc::Parms+1, in_adr); 890 } else { 891 // When slow path is called with a null address, the healed oop will not be written back 892 call->init_req(TypeFunc::Parms+1, igvn.zerocon(T_OBJECT)); 893 } 894 call = igvn.transform(call); 895 896 Node* ctrl = igvn.transform(new ProjNode(call, TypeFunc::Control)); 897 res = igvn.transform(new ProjNode(call, TypeFunc::Parms)); 898 res = igvn.transform(new CheckCastPPNode(ctrl, res, in_val_maybe_null_t)); 899 900 result_region->set_req(2, ctrl); 901 result_val->set_req(2, res); 902 903 result_region = igvn.transform(result_region); 904 result_val = igvn.transform(result_val); 905 906 if (out_ctrl != NULL) { // Added if cond 907 igvn.replace_node(out_ctrl, result_region); 908 } 909 igvn.replace_node(out_res, result_val); 910 } 911 912 // Optimized, low spill, loadbarrier variant using stub specialized on register used 913 void ZBarrierSetC2::expand_loadbarrier_optimized(PhaseMacroExpand* phase, LoadBarrierNode *barrier) const { 914 PhaseIterGVN &igvn = phase->igvn(); 915 #ifdef PRINT_NODE_TRAVERSALS 916 Node* preceding_barrier_node = barrier->in(LoadBarrierNode::Oop); 917 #endif 918 919 Node* in_ctrl = barrier->in(LoadBarrierNode::Control); 920 Node* in_mem = barrier->in(LoadBarrierNode::Memory); 921 Node* in_val = barrier->in(LoadBarrierNode::Oop); 922 Node* in_adr = barrier->in(LoadBarrierNode::Address); 923 924 Node* out_ctrl = barrier->proj_out(LoadBarrierNode::Control); 925 Node* out_res = barrier->proj_out(LoadBarrierNode::Oop); 926 927 assert(barrier->in(LoadBarrierNode::Oop) != NULL, "oop to loadbarrier node cannot be null"); 928 929 #ifdef PRINT_NODE_TRAVERSALS 930 tty->print("\n\n\nBefore barrier optimization:\n"); 931 traverse(barrier, out_ctrl, out_res, -1); 932 933 tty->print("\nBefore barrier optimization: preceding_barrier_node\n"); 934 traverse(preceding_barrier_node, out_ctrl, out_res, -1); 935 #endif 936 937 float unlikely = PROB_UNLIKELY(0.999); 938 939 Node* jthread = igvn.transform(new ThreadLocalNode()); 940 Node* adr = phase->basic_plus_adr(jthread, in_bytes(ZThreadLocalData::address_bad_mask_offset())); 941 Node* bad_mask = igvn.transform(LoadNode::make(igvn, in_ctrl, in_mem, adr, 942 TypeRawPtr::BOTTOM, TypeX_X, TypeX_X->basic_type(), 943 MemNode::unordered)); 944 Node* cast = igvn.transform(new CastP2XNode(in_ctrl, in_val)); 945 Node* obj_masked = igvn.transform(new AndXNode(cast, bad_mask)); 946 Node* cmp = igvn.transform(new CmpXNode(obj_masked, igvn.zerocon(TypeX_X->basic_type()))); 947 Node *bol = igvn.transform(new BoolNode(cmp, BoolTest::ne))->as_Bool(); 948 IfNode* iff = igvn.transform(new IfNode(in_ctrl, bol, unlikely, COUNT_UNKNOWN))->as_If(); 949 Node* then = igvn.transform(new IfTrueNode(iff)); 950 Node* elsen = igvn.transform(new IfFalseNode(iff)); 951 952 Node* slow_path_surrogate; 953 if (!barrier->is_weak()) { 954 slow_path_surrogate = igvn.transform(new LoadBarrierSlowRegNode(then, in_mem, in_adr, in_val->adr_type(), 955 (const TypePtr*) in_val->bottom_type(), MemNode::unordered)); 956 } else { 957 slow_path_surrogate = igvn.transform(new LoadBarrierWeakSlowRegNode(then, in_mem, in_adr, in_val->adr_type(), 958 (const TypePtr*) in_val->bottom_type(), MemNode::unordered)); 959 } 960 961 Node *new_loadp; 962 new_loadp = slow_path_surrogate; 963 // Create the final region/phi pair to converge cntl/data paths to downstream code 964 Node* result_region = igvn.transform(new RegionNode(3)); 965 result_region->set_req(1, then); 966 result_region->set_req(2, elsen); 967 968 Node* result_phi = igvn.transform(new PhiNode(result_region, TypeInstPtr::BOTTOM)); 969 result_phi->set_req(1, new_loadp); 970 result_phi->set_req(2, barrier->in(LoadBarrierNode::Oop)); 971 972 // Finally, connect the original outputs to the barrier region and phi to complete the expansion/substitution 973 // igvn.replace_node(out_ctrl, result_region); 974 if (out_ctrl != NULL) { // added if cond 975 igvn.replace_node(out_ctrl, result_region); 976 } 977 igvn.replace_node(out_res, result_phi); 978 979 assert(barrier->outcnt() == 0,"LoadBarrier macro node has non-null outputs after expansion!"); 980 981 #ifdef PRINT_NODE_TRAVERSALS 982 tty->print("\nAfter barrier optimization: old out_ctrl\n"); 983 traverse(out_ctrl, out_ctrl, out_res, -1); 984 tty->print("\nAfter barrier optimization: old out_res\n"); 985 traverse(out_res, out_ctrl, out_res, -1); 986 tty->print("\nAfter barrier optimization: old barrier\n"); 987 traverse(barrier, out_ctrl, out_res, -1); 988 tty->print("\nAfter barrier optimization: preceding_barrier_node\n"); 989 traverse(preceding_barrier_node, result_region, result_phi, -1); 990 #endif 991 992 assert(is_gc_barrier_node(result_phi), "sanity"); 993 assert(step_over_gc_barrier(result_phi) == in_val, "sanity"); 994 995 return; 996 } 997 998 bool ZBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { 999 Compile* C = Compile::current(); 1000 PhaseIterGVN &igvn = macro->igvn(); 1001 ZBarrierSetC2State* s = state(); 1002 if (s->load_barrier_count() > 0) { 1003 #ifdef ASSERT 1004 verify_gc_barriers(false); 1005 #endif 1006 igvn.set_delay_transform(true); 1007 int skipped = 0; 1008 while (s->load_barrier_count() > skipped) { 1009 int load_barrier_count = s->load_barrier_count(); 1010 LoadBarrierNode * n = s->load_barrier_node(load_barrier_count-1-skipped); 1011 if (igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) { 1012 // Node is unreachable, so don't try to expand it 1013 s->remove_load_barrier_node(n); 1014 continue; 1015 } 1016 if (!n->can_be_eliminated()) { 1017 skipped++; 1018 continue; 1019 } 1020 expand_loadbarrier_node(macro, n); 1021 assert(s->load_barrier_count() < load_barrier_count, "must have deleted a node from load barrier list"); 1022 if (C->failing()) return true; 1023 } 1024 while (s->load_barrier_count() > 0) { 1025 int load_barrier_count = s->load_barrier_count(); 1026 LoadBarrierNode* n = s->load_barrier_node(load_barrier_count - 1); 1027 assert(!(igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())), "should have been processed already"); 1028 assert(!n->can_be_eliminated(), "should have been processed already"); 1029 expand_loadbarrier_node(macro, n); 1030 assert(s->load_barrier_count() < load_barrier_count, "must have deleted a node from load barrier list"); 1031 if (C->failing()) return true; 1032 } 1033 igvn.set_delay_transform(false); 1034 igvn.optimize(); 1035 if (C->failing()) return true; 1036 } 1037 return false; 1038 } 1039 1040 // == Loop optimization == 1041 1042 static bool replace_with_dominating_barrier(PhaseIdealLoop* phase, LoadBarrierNode* lb, bool last_round) { 1043 PhaseIterGVN &igvn = phase->igvn(); 1044 Compile* C = Compile::current(); 1045 1046 LoadBarrierNode* lb2 = lb->has_dominating_barrier(phase, false, last_round); 1047 if (lb2 != NULL) { 1048 if (lb->in(LoadBarrierNode::Oop) != lb2->in(LoadBarrierNode::Oop)) { 1049 assert(lb->in(LoadBarrierNode::Address) == lb2->in(LoadBarrierNode::Address), ""); 1050 igvn.replace_input_of(lb, LoadBarrierNode::Similar, lb2->proj_out(LoadBarrierNode::Oop)); 1051 C->set_major_progress(); 1052 } else { 1053 // That transformation may cause the Similar edge on dominated load barriers to be invalid 1054 lb->fix_similar_in_uses(&igvn); 1055 1056 Node* val = lb->proj_out(LoadBarrierNode::Oop); 1057 assert(lb2->has_true_uses(), ""); 1058 assert(lb2->in(LoadBarrierNode::Oop) == lb->in(LoadBarrierNode::Oop), ""); 1059 1060 phase->lazy_update(lb, lb->in(LoadBarrierNode::Control)); 1061 phase->lazy_replace(lb->proj_out(LoadBarrierNode::Control), lb->in(LoadBarrierNode::Control)); 1062 igvn.replace_node(val, lb2->proj_out(LoadBarrierNode::Oop)); 1063 1064 return true; 1065 } 1066 } 1067 return false; 1068 } 1069 1070 static Node* find_dominating_memory(PhaseIdealLoop* phase, Node* mem, Node* dom, int i) { 1071 assert(dom->is_Region() || i == -1, ""); 1072 Node* m = mem; 1073 while(phase->is_dominator(dom, phase->has_ctrl(m) ? phase->get_ctrl(m) : m->in(0))) { 1074 if (m->is_Mem()) { 1075 assert(m->as_Mem()->adr_type() == TypeRawPtr::BOTTOM, ""); 1076 m = m->in(MemNode::Memory); 1077 } else if (m->is_MergeMem()) { 1078 m = m->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1079 } else if (m->is_Phi()) { 1080 if (m->in(0) == dom && i != -1) { 1081 m = m->in(i); 1082 break; 1083 } else { 1084 m = m->in(LoopNode::EntryControl); 1085 } 1086 } else if (m->is_Proj()) { 1087 m = m->in(0); 1088 } else if (m->is_SafePoint() || m->is_MemBar()) { 1089 m = m->in(TypeFunc::Memory); 1090 } else { 1091 #ifdef ASSERT 1092 m->dump(); 1093 #endif 1094 ShouldNotReachHere(); 1095 } 1096 } 1097 return m; 1098 } 1099 1100 static LoadBarrierNode* clone_load_barrier(PhaseIdealLoop* phase, LoadBarrierNode* lb, Node* ctl, Node* mem, Node* oop_in) { 1101 PhaseIterGVN &igvn = phase->igvn(); 1102 Compile* C = Compile::current(); 1103 Node* the_clone = lb->clone(); 1104 the_clone->set_req(LoadBarrierNode::Control, ctl); 1105 the_clone->set_req(LoadBarrierNode::Memory, mem); 1106 if (oop_in != NULL) { 1107 the_clone->set_req(LoadBarrierNode::Oop, oop_in); 1108 } 1109 1110 LoadBarrierNode* new_lb = the_clone->as_LoadBarrier(); 1111 igvn.register_new_node_with_optimizer(new_lb); 1112 IdealLoopTree *loop = phase->get_loop(new_lb->in(0)); 1113 phase->set_ctrl(new_lb, new_lb->in(0)); 1114 phase->set_loop(new_lb, loop); 1115 phase->set_idom(new_lb, new_lb->in(0), phase->dom_depth(new_lb->in(0))+1); 1116 if (!loop->_child) { 1117 loop->_body.push(new_lb); 1118 } 1119 1120 Node* proj_ctl = new ProjNode(new_lb, LoadBarrierNode::Control); 1121 igvn.register_new_node_with_optimizer(proj_ctl); 1122 phase->set_ctrl(proj_ctl, proj_ctl->in(0)); 1123 phase->set_loop(proj_ctl, loop); 1124 phase->set_idom(proj_ctl, new_lb, phase->dom_depth(new_lb)+1); 1125 if (!loop->_child) { 1126 loop->_body.push(proj_ctl); 1127 } 1128 1129 Node* proj_oop = new ProjNode(new_lb, LoadBarrierNode::Oop); 1130 phase->register_new_node(proj_oop, new_lb); 1131 1132 if (!new_lb->in(LoadBarrierNode::Similar)->is_top()) { 1133 LoadBarrierNode* similar = new_lb->in(LoadBarrierNode::Similar)->in(0)->as_LoadBarrier(); 1134 if (!phase->is_dominator(similar, ctl)) { 1135 igvn.replace_input_of(new_lb, LoadBarrierNode::Similar, C->top()); 1136 } 1137 } 1138 1139 return new_lb; 1140 } 1141 1142 static void replace_barrier(PhaseIdealLoop* phase, LoadBarrierNode* lb, Node* new_val) { 1143 PhaseIterGVN &igvn = phase->igvn(); 1144 Node* val = lb->proj_out(LoadBarrierNode::Oop); 1145 igvn.replace_node(val, new_val); 1146 phase->lazy_update(lb, lb->in(LoadBarrierNode::Control)); 1147 phase->lazy_replace(lb->proj_out(LoadBarrierNode::Control), lb->in(LoadBarrierNode::Control)); 1148 } 1149 1150 static bool split_barrier_thru_phi(PhaseIdealLoop* phase, LoadBarrierNode* lb) { 1151 PhaseIterGVN &igvn = phase->igvn(); 1152 Compile* C = Compile::current(); 1153 1154 if (lb->in(LoadBarrierNode::Oop)->is_Phi()) { 1155 Node* oop_phi = lb->in(LoadBarrierNode::Oop); 1156 1157 if (oop_phi->in(2) == oop_phi) { 1158 // Ignore phis with only one input 1159 return false; 1160 } 1161 1162 if (phase->is_dominator(phase->get_ctrl(lb->in(LoadBarrierNode::Address)), 1163 oop_phi->in(0)) && phase->get_ctrl(lb->in(LoadBarrierNode::Address)) != oop_phi->in(0)) { 1164 // That transformation may cause the Similar edge on dominated load barriers to be invalid 1165 lb->fix_similar_in_uses(&igvn); 1166 1167 RegionNode* region = oop_phi->in(0)->as_Region(); 1168 1169 int backedge = LoopNode::LoopBackControl; 1170 if (region->is_Loop() && region->in(backedge)->is_Proj() && region->in(backedge)->in(0)->is_If()) { 1171 Node* c = region->in(backedge)->in(0)->in(0); 1172 assert(c->unique_ctrl_out() == region->in(backedge)->in(0), ""); 1173 Node* oop = lb->in(LoadBarrierNode::Oop)->in(backedge); 1174 Node* oop_c = phase->has_ctrl(oop) ? phase->get_ctrl(oop) : oop; 1175 if (!phase->is_dominator(oop_c, c)) { 1176 return false; 1177 } 1178 } 1179 1180 // If the node on the backedge above the phi is the node itself - we have a self loop. 1181 // Don't clone - this will be folded later. 1182 if (oop_phi->in(LoopNode::LoopBackControl) == lb->proj_out(LoadBarrierNode::Oop)) { 1183 return false; 1184 } 1185 1186 bool is_strip_mined = region->is_CountedLoop() && region->as_CountedLoop()->is_strip_mined(); 1187 Node *phi = oop_phi->clone(); 1188 1189 for (uint i = 1; i < region->req(); i++) { 1190 Node* ctrl = region->in(i); 1191 if (ctrl != C->top()) { 1192 assert(!phase->is_dominator(ctrl, region) || region->is_Loop(), ""); 1193 1194 Node* mem = lb->in(LoadBarrierNode::Memory); 1195 Node* m = find_dominating_memory(phase, mem, region, i); 1196 1197 if (region->is_Loop() && i == LoopNode::LoopBackControl && ctrl->is_Proj() && ctrl->in(0)->is_If()) { 1198 ctrl = ctrl->in(0)->in(0); 1199 } else if (region->is_Loop() && is_strip_mined) { 1200 // If this is a strip mined loop, control must move above OuterStripMinedLoop 1201 assert(i == LoopNode::EntryControl, "check"); 1202 assert(ctrl->is_OuterStripMinedLoop(), "sanity"); 1203 ctrl = ctrl->as_OuterStripMinedLoop()->in(LoopNode::EntryControl); 1204 } 1205 1206 LoadBarrierNode* new_lb = clone_load_barrier(phase, lb, ctrl, m, lb->in(LoadBarrierNode::Oop)->in(i)); 1207 Node* out_ctrl = new_lb->proj_out(LoadBarrierNode::Control); 1208 1209 if (is_strip_mined && (i == LoopNode::EntryControl)) { 1210 assert(region->in(i)->is_OuterStripMinedLoop(), ""); 1211 igvn.replace_input_of(region->in(i), i, out_ctrl); 1212 phase->set_idom(region->in(i), out_ctrl, phase->dom_depth(out_ctrl)); 1213 } else if (ctrl == region->in(i)) { 1214 igvn.replace_input_of(region, i, out_ctrl); 1215 // Only update the idom if is the loop entry we are updating 1216 // - A loop backedge doesn't change the idom 1217 if (region->is_Loop() && i == LoopNode::EntryControl) { 1218 phase->set_idom(region, out_ctrl, phase->dom_depth(out_ctrl)); 1219 } 1220 } else { 1221 Node* iff = region->in(i)->in(0); 1222 igvn.replace_input_of(iff, 0, out_ctrl); 1223 phase->set_idom(iff, out_ctrl, phase->dom_depth(out_ctrl)+1); 1224 } 1225 phi->set_req(i, new_lb->proj_out(LoadBarrierNode::Oop)); 1226 } 1227 } 1228 phase->register_new_node(phi, region); 1229 replace_barrier(phase, lb, phi); 1230 1231 if (region->is_Loop()) { 1232 // Load barrier moved to the back edge of the Loop may now 1233 // have a safepoint on the path to the barrier on the Similar 1234 // edge 1235 igvn.replace_input_of(phi->in(LoopNode::LoopBackControl)->in(0), LoadBarrierNode::Similar, C->top()); 1236 Node* head = region->in(LoopNode::EntryControl); 1237 phase->set_idom(region, head, phase->dom_depth(head)+1); 1238 phase->recompute_dom_depth(); 1239 if (head->is_CountedLoop() && head->as_CountedLoop()->is_main_loop()) { 1240 head->as_CountedLoop()->set_normal_loop(); 1241 } 1242 } 1243 1244 return true; 1245 } 1246 } 1247 1248 return false; 1249 } 1250 1251 static bool move_out_of_loop(PhaseIdealLoop* phase, LoadBarrierNode* lb) { 1252 PhaseIterGVN &igvn = phase->igvn(); 1253 IdealLoopTree *lb_loop = phase->get_loop(lb->in(0)); 1254 if (lb_loop != phase->ltree_root() && !lb_loop->_irreducible) { 1255 Node* oop_ctrl = phase->get_ctrl(lb->in(LoadBarrierNode::Oop)); 1256 IdealLoopTree *oop_loop = phase->get_loop(oop_ctrl); 1257 IdealLoopTree* adr_loop = phase->get_loop(phase->get_ctrl(lb->in(LoadBarrierNode::Address))); 1258 if (!lb_loop->is_member(oop_loop) && !lb_loop->is_member(adr_loop)) { 1259 // That transformation may cause the Similar edge on dominated load barriers to be invalid 1260 lb->fix_similar_in_uses(&igvn); 1261 1262 Node* head = lb_loop->_head; 1263 assert(head->is_Loop(), ""); 1264 1265 if (phase->is_dominator(head, oop_ctrl)) { 1266 assert(oop_ctrl->Opcode() == Op_CProj && oop_ctrl->in(0)->Opcode() == Op_NeverBranch, ""); 1267 assert(lb_loop->is_member(phase->get_loop(oop_ctrl->in(0)->in(0))), ""); 1268 return false; 1269 } 1270 1271 if (head->is_CountedLoop()) { 1272 CountedLoopNode* cloop = head->as_CountedLoop(); 1273 if (cloop->is_main_loop()) { 1274 cloop->set_normal_loop(); 1275 } 1276 // When we are moving barrier out of a counted loop, 1277 // make sure we move it all the way out of the strip mined outer loop. 1278 if (cloop->is_strip_mined()) { 1279 head = cloop->outer_loop(); 1280 } 1281 } 1282 1283 Node* mem = lb->in(LoadBarrierNode::Memory); 1284 Node* m = find_dominating_memory(phase, mem, head, -1); 1285 1286 LoadBarrierNode* new_lb = clone_load_barrier(phase, lb, head->in(LoopNode::EntryControl), m, NULL); 1287 1288 assert(phase->idom(head) == head->in(LoopNode::EntryControl), ""); 1289 Node* proj_ctl = new_lb->proj_out(LoadBarrierNode::Control); 1290 igvn.replace_input_of(head, LoopNode::EntryControl, proj_ctl); 1291 phase->set_idom(head, proj_ctl, phase->dom_depth(proj_ctl) + 1); 1292 1293 replace_barrier(phase, lb, new_lb->proj_out(LoadBarrierNode::Oop)); 1294 1295 phase->recompute_dom_depth(); 1296 1297 return true; 1298 } 1299 } 1300 1301 return false; 1302 } 1303 1304 static bool common_barriers(PhaseIdealLoop* phase, LoadBarrierNode* lb) { 1305 PhaseIterGVN &igvn = phase->igvn(); 1306 Node* in_val = lb->in(LoadBarrierNode::Oop); 1307 for (DUIterator_Fast imax, i = in_val->fast_outs(imax); i < imax; i++) { 1308 Node* u = in_val->fast_out(i); 1309 if (u != lb && u->is_LoadBarrier() && u->as_LoadBarrier()->has_true_uses()) { 1310 Node* this_ctrl = lb->in(LoadBarrierNode::Control); 1311 Node* other_ctrl = u->in(LoadBarrierNode::Control); 1312 1313 Node* lca = phase->dom_lca(this_ctrl, other_ctrl); 1314 bool ok = true; 1315 1316 Node* proj1 = NULL; 1317 Node* proj2 = NULL; 1318 1319 while (this_ctrl != lca && ok) { 1320 if (this_ctrl->in(0) != NULL && 1321 this_ctrl->in(0)->is_MultiBranch()) { 1322 if (this_ctrl->in(0)->in(0) == lca) { 1323 assert(proj1 == NULL, ""); 1324 assert(this_ctrl->is_Proj(), ""); 1325 proj1 = this_ctrl; 1326 } else if (!(this_ctrl->in(0)->is_If() && this_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none))) { 1327 ok = false; 1328 } 1329 } 1330 this_ctrl = phase->idom(this_ctrl); 1331 } 1332 while (other_ctrl != lca && ok) { 1333 if (other_ctrl->in(0) != NULL && 1334 other_ctrl->in(0)->is_MultiBranch()) { 1335 if (other_ctrl->in(0)->in(0) == lca) { 1336 assert(other_ctrl->is_Proj(), ""); 1337 assert(proj2 == NULL, ""); 1338 proj2 = other_ctrl; 1339 } else if (!(other_ctrl->in(0)->is_If() && other_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none))) { 1340 ok = false; 1341 } 1342 } 1343 other_ctrl = phase->idom(other_ctrl); 1344 } 1345 assert(proj1 == NULL || proj2 == NULL || proj1->in(0) == proj2->in(0), ""); 1346 if (ok && proj1 && proj2 && proj1 != proj2 && proj1->in(0)->is_If()) { 1347 // That transformation may cause the Similar edge on dominated load barriers to be invalid 1348 lb->fix_similar_in_uses(&igvn); 1349 u->as_LoadBarrier()->fix_similar_in_uses(&igvn); 1350 1351 Node* split = lca->unique_ctrl_out(); 1352 assert(split->in(0) == lca, ""); 1353 1354 Node* mem = lb->in(LoadBarrierNode::Memory); 1355 Node* m = find_dominating_memory(phase, mem, split, -1); 1356 LoadBarrierNode* new_lb = clone_load_barrier(phase, lb, lca, m, NULL); 1357 1358 Node* proj_ctl = new_lb->proj_out(LoadBarrierNode::Control); 1359 igvn.replace_input_of(split, 0, new_lb->proj_out(LoadBarrierNode::Control)); 1360 phase->set_idom(split, proj_ctl, phase->dom_depth(proj_ctl)+1); 1361 1362 Node* proj_oop = new_lb->proj_out(LoadBarrierNode::Oop); 1363 replace_barrier(phase, lb, proj_oop); 1364 replace_barrier(phase, u->as_LoadBarrier(), proj_oop); 1365 1366 phase->recompute_dom_depth(); 1367 1368 return true; 1369 } 1370 } 1371 } 1372 1373 return false; 1374 } 1375 1376 static void optimize_load_barrier(PhaseIdealLoop* phase, LoadBarrierNode* lb, bool last_round) { 1377 Compile* C = Compile::current(); 1378 1379 if (!C->directive()->ZOptimizeLoadBarriersOption) { 1380 return; 1381 } 1382 1383 if (lb->has_true_uses()) { 1384 if (replace_with_dominating_barrier(phase, lb, last_round)) { 1385 return; 1386 } 1387 1388 if (split_barrier_thru_phi(phase, lb)) { 1389 return; 1390 } 1391 1392 if (move_out_of_loop(phase, lb)) { 1393 return; 1394 } 1395 1396 if (common_barriers(phase, lb)) { 1397 return; 1398 } 1399 } 1400 } 1401 1402 void ZBarrierSetC2::loop_optimize_gc_barrier(PhaseIdealLoop* phase, Node* node, bool last_round) { 1403 if (node->is_LoadBarrier()) { 1404 optimize_load_barrier(phase, node->as_LoadBarrier(), last_round); 1405 } 1406 } 1407 1408 Node* ZBarrierSetC2::step_over_gc_barrier(Node* c) const { 1409 Node* node = c; 1410 1411 // 1. This step follows potential oop projections of a load barrier before expansion 1412 if (node->is_Proj()) { 1413 node = node->in(0); 1414 } 1415 1416 // 2. This step checks for unexpanded load barriers 1417 if (node->is_LoadBarrier()) { 1418 return node->in(LoadBarrierNode::Oop); 1419 } 1420 1421 // 3. This step checks for the phi corresponding to an optimized load barrier expansion 1422 if (node->is_Phi()) { 1423 PhiNode* phi = node->as_Phi(); 1424 Node* n = phi->in(1); 1425 if (n != NULL && (n->is_LoadBarrierSlowReg() || n->is_LoadBarrierWeakSlowReg())) { 1426 assert(c == node, "projections from step 1 should only be seen before macro expansion"); 1427 return phi->in(2); 1428 } 1429 } 1430 1431 return c; 1432 } 1433 1434 bool ZBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, ArrayCopyPhase phase) const { 1435 return type == T_OBJECT || type == T_ARRAY; 1436 } 1437 1438 bool ZBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode) const { 1439 bool handled; 1440 switch (opcode) { 1441 case Op_LoadBarrierSlowReg: 1442 case Op_LoadBarrierWeakSlowReg: 1443 #ifdef ASSERT 1444 if (VerifyOptoOopOffsets) { 1445 MemNode* mem = n->as_Mem(); 1446 // Check to see if address types have grounded out somehow. 1447 const TypeInstPtr* tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 1448 ciInstanceKlass* k = tp->klass()->as_instance_klass(); 1449 bool oop_offset_is_sane = k->contains_field_offset(tp->offset()); 1450 assert(!tp || oop_offset_is_sane, ""); 1451 } 1452 #endif 1453 handled = true; 1454 break; 1455 default: 1456 handled = false; 1457 } 1458 return handled; 1459 } 1460 1461 // == Verification == 1462 1463 #ifdef ASSERT 1464 1465 static bool look_for_barrier(Node* n, bool post_parse, VectorSet& visited) { 1466 if (visited.test_set(n->_idx)) { 1467 return true; 1468 } 1469 1470 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1471 Node* u = n->fast_out(i); 1472 if (u->is_LoadBarrier()) { 1473 } else if ((u->is_Phi() || u->is_CMove()) && !post_parse) { 1474 if (!look_for_barrier(u, post_parse, visited)) { 1475 return false; 1476 } 1477 } else if (u->Opcode() == Op_EncodeP || u->Opcode() == Op_DecodeN) { 1478 if (!look_for_barrier(u, post_parse, visited)) { 1479 return false; 1480 } 1481 } else if (u->Opcode() != Op_SCMemProj) { 1482 tty->print("bad use"); u->dump(); 1483 return false; 1484 } 1485 } 1486 1487 return true; 1488 } 1489 1490 void ZBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const { 1491 if (phase == BarrierSetC2::BeforeCodeGen) return; 1492 bool post_parse = phase == BarrierSetC2::BeforeOptimize; 1493 verify_gc_barriers(post_parse); 1494 } 1495 1496 void ZBarrierSetC2::verify_gc_barriers(bool post_parse) const { 1497 ZBarrierSetC2State* s = state(); 1498 Compile* C = Compile::current(); 1499 ResourceMark rm; 1500 VectorSet visited(Thread::current()->resource_area()); 1501 for (int i = 0; i < s->load_barrier_count(); i++) { 1502 LoadBarrierNode* n = s->load_barrier_node(i); 1503 1504 // The dominating barrier on the same address if it exists and 1505 // this barrier must not be applied on the value from the same 1506 // load otherwise the value is not reloaded before it's used the 1507 // second time. 1508 assert(n->in(LoadBarrierNode::Similar)->is_top() || 1509 (n->in(LoadBarrierNode::Similar)->in(0)->is_LoadBarrier() && 1510 n->in(LoadBarrierNode::Similar)->in(0)->in(LoadBarrierNode::Address) == n->in(LoadBarrierNode::Address) && 1511 n->in(LoadBarrierNode::Similar)->in(0)->in(LoadBarrierNode::Oop) != n->in(LoadBarrierNode::Oop)), 1512 "broken similar edge"); 1513 1514 assert(post_parse || n->as_LoadBarrier()->has_true_uses(), 1515 "found unneeded load barrier"); 1516 1517 // Several load barrier nodes chained through their Similar edge 1518 // break the code that remove the barriers in final graph reshape. 1519 assert(n->in(LoadBarrierNode::Similar)->is_top() || 1520 (n->in(LoadBarrierNode::Similar)->in(0)->is_LoadBarrier() && 1521 n->in(LoadBarrierNode::Similar)->in(0)->in(LoadBarrierNode::Similar)->is_top()), 1522 "chain of Similar load barriers"); 1523 1524 if (!n->in(LoadBarrierNode::Similar)->is_top()) { 1525 ResourceMark rm; 1526 Unique_Node_List wq; 1527 Node* other = n->in(LoadBarrierNode::Similar)->in(0); 1528 wq.push(n); 1529 bool ok = true; 1530 bool dom_found = false; 1531 for (uint next = 0; next < wq.size(); ++next) { 1532 Node *n = wq.at(next); 1533 assert(n->is_CFG(), ""); 1534 assert(!n->is_SafePoint(), ""); 1535 1536 if (n == other) { 1537 continue; 1538 } 1539 1540 if (n->is_Region()) { 1541 for (uint i = 1; i < n->req(); i++) { 1542 Node* m = n->in(i); 1543 if (m != NULL) { 1544 wq.push(m); 1545 } 1546 } 1547 } else { 1548 Node* m = n->in(0); 1549 if (m != NULL) { 1550 wq.push(m); 1551 } 1552 } 1553 } 1554 } 1555 1556 if (ZVerifyLoadBarriers) { 1557 if ((n->is_Load() || n->is_LoadStore()) && n->bottom_type()->make_oopptr() != NULL) { 1558 visited.Clear(); 1559 bool found = look_for_barrier(n, post_parse, visited); 1560 if (!found) { 1561 n->dump(1); 1562 n->dump(-3); 1563 stringStream ss; 1564 C->method()->print_short_name(&ss); 1565 tty->print_cr("-%s-", ss.as_string()); 1566 assert(found, ""); 1567 } 1568 } 1569 } 1570 } 1571 } 1572 1573 #endif