1 /* 2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved. 3 * Copyright (c) 2024, Alibaba Group Holding Limited. All rights reserved. 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 5 * 6 * This code is free software; you can redistribute it and/or modify it 7 * under the terms of the GNU General Public License version 2 only, as 8 * published by the Free Software Foundation. 9 * 10 * This code is distributed in the hope that it will be useful, but WITHOUT 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 13 * version 2 for more details (a copy is included in the LICENSE file that 14 * accompanied this code). 15 * 16 * You should have received a copy of the GNU General Public License version 17 * 2 along with this work; if not, write to the Free Software Foundation, 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 19 * 20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 21 * or visit www.oracle.com if you need additional information or have any 22 * questions. 23 * 24 */ 25 26 #include "classfile/javaClasses.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "gc/shared/barrierSet.hpp" 29 #include "gc/shared/c2/barrierSetC2.hpp" 30 #include "gc/shared/tlab_globals.hpp" 31 #include "memory/allocation.inline.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "oops/objArrayKlass.hpp" 34 #include "opto/addnode.hpp" 35 #include "opto/arraycopynode.hpp" 36 #include "opto/cfgnode.hpp" 37 #include "opto/regalloc.hpp" 38 #include "opto/compile.hpp" 39 #include "opto/connode.hpp" 40 #include "opto/convertnode.hpp" 41 #include "opto/loopnode.hpp" 42 #include "opto/machnode.hpp" 43 #include "opto/matcher.hpp" 44 #include "opto/memnode.hpp" 45 #include "opto/mempointer.hpp" 46 #include "opto/mulnode.hpp" 47 #include "opto/narrowptrnode.hpp" 48 #include "opto/phaseX.hpp" 49 #include "opto/regmask.hpp" 50 #include "opto/rootnode.hpp" 51 #include "opto/traceMergeStoresTag.hpp" 52 #include "opto/vectornode.hpp" 53 #include "utilities/align.hpp" 54 #include "utilities/copy.hpp" 55 #include "utilities/macros.hpp" 56 #include "utilities/powerOfTwo.hpp" 57 #include "utilities/vmError.hpp" 58 59 // Portions of code courtesy of Clifford Click 60 61 // Optimization - Graph Style 62 63 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 64 65 //============================================================================= 66 uint MemNode::size_of() const { return sizeof(*this); } 67 68 const TypePtr *MemNode::adr_type() const { 69 Node* adr = in(Address); 70 if (adr == nullptr) return nullptr; // node is dead 71 const TypePtr* cross_check = nullptr; 72 DEBUG_ONLY(cross_check = _adr_type); 73 return calculate_adr_type(adr->bottom_type(), cross_check); 74 } 75 76 bool MemNode::check_if_adr_maybe_raw(Node* adr) { 77 if (adr != nullptr) { 78 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) { 79 return true; 80 } 81 } 82 return false; 83 } 84 85 #ifndef PRODUCT 86 void MemNode::dump_spec(outputStream *st) const { 87 if (in(Address) == nullptr) return; // node is dead 88 #ifndef ASSERT 89 // fake the missing field 90 const TypePtr* _adr_type = nullptr; 91 if (in(Address) != nullptr) 92 _adr_type = in(Address)->bottom_type()->isa_ptr(); 93 #endif 94 dump_adr_type(this, _adr_type, st); 95 96 Compile* C = Compile::current(); 97 if (C->alias_type(_adr_type)->is_volatile()) { 98 st->print(" Volatile!"); 99 } 100 if (_unaligned_access) { 101 st->print(" unaligned"); 102 } 103 if (_mismatched_access) { 104 st->print(" mismatched"); 105 } 106 if (_unsafe_access) { 107 st->print(" unsafe"); 108 } 109 } 110 111 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 112 st->print(" @"); 113 if (adr_type == nullptr) { 114 st->print("null"); 115 } else { 116 adr_type->dump_on(st); 117 Compile* C = Compile::current(); 118 Compile::AliasType* atp = nullptr; 119 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 120 if (atp == nullptr) 121 st->print(", idx=?\?;"); 122 else if (atp->index() == Compile::AliasIdxBot) 123 st->print(", idx=Bot;"); 124 else if (atp->index() == Compile::AliasIdxTop) 125 st->print(", idx=Top;"); 126 else if (atp->index() == Compile::AliasIdxRaw) 127 st->print(", idx=Raw;"); 128 else { 129 ciField* field = atp->field(); 130 if (field) { 131 st->print(", name="); 132 field->print_name_on(st); 133 } 134 st->print(", idx=%d;", atp->index()); 135 } 136 } 137 } 138 139 extern void print_alias_types(); 140 141 #endif 142 143 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 144 assert((t_oop != nullptr), "sanity"); 145 bool is_instance = t_oop->is_known_instance_field(); 146 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 147 (load != nullptr) && load->is_Load() && 148 (phase->is_IterGVN() != nullptr); 149 if (!(is_instance || is_boxed_value_load)) 150 return mchain; // don't try to optimize non-instance types 151 uint instance_id = t_oop->instance_id(); 152 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory); 153 Node *prev = nullptr; 154 Node *result = mchain; 155 while (prev != result) { 156 prev = result; 157 if (result == start_mem) 158 break; // hit one of our sentinels 159 // skip over a call which does not affect this memory slice 160 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 161 Node *proj_in = result->in(0); 162 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 163 break; // hit one of our sentinels 164 } else if (proj_in->is_Call()) { 165 // ArrayCopyNodes processed here as well 166 CallNode *call = proj_in->as_Call(); 167 if (!call->may_modify(t_oop, phase)) { // returns false for instances 168 result = call->in(TypeFunc::Memory); 169 } 170 } else if (proj_in->is_Initialize()) { 171 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 172 // Stop if this is the initialization for the object instance which 173 // contains this memory slice, otherwise skip over it. 174 if ((alloc == nullptr) || (alloc->_idx == instance_id)) { 175 break; 176 } 177 if (is_instance) { 178 result = proj_in->in(TypeFunc::Memory); 179 } else if (is_boxed_value_load) { 180 Node* klass = alloc->in(AllocateNode::KlassNode); 181 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 182 if (tklass->klass_is_exact() && !tklass->exact_klass()->equals(t_oop->is_instptr()->exact_klass())) { 183 result = proj_in->in(TypeFunc::Memory); // not related allocation 184 } 185 } 186 } else if (proj_in->is_MemBar()) { 187 ArrayCopyNode* ac = nullptr; 188 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) { 189 break; 190 } 191 result = proj_in->in(TypeFunc::Memory); 192 } else if (proj_in->is_top()) { 193 break; // dead code 194 } else { 195 assert(false, "unexpected projection"); 196 } 197 } else if (result->is_ClearArray()) { 198 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 199 // Can not bypass initialization of the instance 200 // we are looking for. 201 break; 202 } 203 // Otherwise skip it (the call updated 'result' value). 204 } else if (result->is_MergeMem()) { 205 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, nullptr, tty); 206 } 207 } 208 return result; 209 } 210 211 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 212 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 213 if (t_oop == nullptr) 214 return mchain; // don't try to optimize non-oop types 215 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 216 bool is_instance = t_oop->is_known_instance_field(); 217 PhaseIterGVN *igvn = phase->is_IterGVN(); 218 if (is_instance && igvn != nullptr && result->is_Phi()) { 219 PhiNode *mphi = result->as_Phi(); 220 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 221 const TypePtr *t = mphi->adr_type(); 222 bool do_split = false; 223 // In the following cases, Load memory input can be further optimized based on 224 // its precise address type 225 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) { 226 do_split = true; 227 } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) { 228 const TypeOopPtr* mem_t = 229 t->is_oopptr()->cast_to_exactness(true) 230 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 231 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()); 232 if (t_oop->isa_aryptr()) { 233 mem_t = mem_t->is_aryptr() 234 ->cast_to_stable(t_oop->is_aryptr()->is_stable()) 235 ->cast_to_size(t_oop->is_aryptr()->size()) 236 ->with_offset(t_oop->is_aryptr()->offset()) 237 ->is_aryptr(); 238 } 239 do_split = mem_t == t_oop; 240 } 241 if (do_split) { 242 // clone the Phi with our address type 243 result = mphi->split_out_instance(t_adr, igvn); 244 } else { 245 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 246 } 247 } 248 return result; 249 } 250 251 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 252 uint alias_idx = phase->C->get_alias_index(tp); 253 Node *mem = mmem; 254 #ifdef ASSERT 255 { 256 // Check that current type is consistent with the alias index used during graph construction 257 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 258 bool consistent = adr_check == nullptr || adr_check->empty() || 259 phase->C->must_alias(adr_check, alias_idx ); 260 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 261 if( !consistent && adr_check != nullptr && !adr_check->empty() && 262 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 263 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 264 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 265 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 266 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 267 // don't assert if it is dead code. 268 consistent = true; 269 } 270 if( !consistent ) { 271 st->print("alias_idx==%d, adr_check==", alias_idx); 272 if( adr_check == nullptr ) { 273 st->print("null"); 274 } else { 275 adr_check->dump(); 276 } 277 st->cr(); 278 print_alias_types(); 279 assert(consistent, "adr_check must match alias idx"); 280 } 281 } 282 #endif 283 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 284 // means an array I have not precisely typed yet. Do not do any 285 // alias stuff with it any time soon. 286 const TypeOopPtr *toop = tp->isa_oopptr(); 287 if (tp->base() != Type::AnyPtr && 288 !(toop && 289 toop->isa_instptr() && 290 toop->is_instptr()->instance_klass()->is_java_lang_Object() && 291 toop->offset() == Type::OffsetBot)) { 292 // IGVN _delay_transform may be set to true and if that is the case and mmem 293 // is already a registered node then the validation inside transform will 294 // complain. 295 Node* m = mmem; 296 PhaseIterGVN* igvn = phase->is_IterGVN(); 297 if (igvn == nullptr || !igvn->delay_transform()) { 298 // compress paths and change unreachable cycles to TOP 299 // If not, we can update the input infinitely along a MergeMem cycle 300 // Equivalent code in PhiNode::Ideal 301 m = phase->transform(mmem); 302 } 303 // If transformed to a MergeMem, get the desired slice 304 // Otherwise the returned node represents memory for every slice 305 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 306 // Update input if it is progress over what we have now 307 } 308 return mem; 309 } 310 311 //--------------------------Ideal_common--------------------------------------- 312 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 313 // Unhook non-raw memories from complete (macro-expanded) initializations. 314 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 315 // If our control input is a dead region, kill all below the region 316 Node *ctl = in(MemNode::Control); 317 if (ctl && remove_dead_region(phase, can_reshape)) 318 return this; 319 ctl = in(MemNode::Control); 320 // Don't bother trying to transform a dead node 321 if (ctl && ctl->is_top()) return NodeSentinel; 322 323 PhaseIterGVN *igvn = phase->is_IterGVN(); 324 // Wait if control on the worklist. 325 if (ctl && can_reshape && igvn != nullptr) { 326 Node* bol = nullptr; 327 Node* cmp = nullptr; 328 if (ctl->in(0)->is_If()) { 329 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 330 bol = ctl->in(0)->in(1); 331 if (bol->is_Bool()) 332 cmp = ctl->in(0)->in(1)->in(1); 333 } 334 if (igvn->_worklist.member(ctl) || 335 (bol != nullptr && igvn->_worklist.member(bol)) || 336 (cmp != nullptr && igvn->_worklist.member(cmp)) ) { 337 // This control path may be dead. 338 // Delay this memory node transformation until the control is processed. 339 igvn->_worklist.push(this); 340 return NodeSentinel; // caller will return null 341 } 342 } 343 // Ignore if memory is dead, or self-loop 344 Node *mem = in(MemNode::Memory); 345 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null 346 assert(mem != this, "dead loop in MemNode::Ideal"); 347 348 if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) { 349 // This memory slice may be dead. 350 // Delay this mem node transformation until the memory is processed. 351 igvn->_worklist.push(this); 352 return NodeSentinel; // caller will return null 353 } 354 355 Node *address = in(MemNode::Address); 356 const Type *t_adr = phase->type(address); 357 if (t_adr == Type::TOP) return NodeSentinel; // caller will return null 358 359 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) { 360 // Unsafe off-heap access with zero address. Remove access and other control users 361 // to not confuse optimizations and add a HaltNode to fail if this is ever executed. 362 assert(ctl != nullptr, "unsafe accesses should be control dependent"); 363 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) { 364 Node* u = ctl->fast_out(i); 365 if (u != ctl) { 366 igvn->rehash_node_delayed(u); 367 int nb = u->replace_edge(ctl, phase->C->top(), igvn); 368 --i, imax -= nb; 369 } 370 } 371 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr)); 372 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address")); 373 phase->C->root()->add_req(halt); 374 return this; 375 } 376 377 if (can_reshape && igvn != nullptr && 378 (igvn->_worklist.member(address) || 379 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) { 380 // The address's base and type may change when the address is processed. 381 // Delay this mem node transformation until the address is processed. 382 igvn->_worklist.push(this); 383 return NodeSentinel; // caller will return null 384 } 385 386 // Do NOT remove or optimize the next lines: ensure a new alias index 387 // is allocated for an oop pointer type before Escape Analysis. 388 // Note: C++ will not remove it since the call has side effect. 389 if (t_adr->isa_oopptr()) { 390 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 391 } 392 393 Node* base = nullptr; 394 if (address->is_AddP()) { 395 base = address->in(AddPNode::Base); 396 } 397 if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 398 !t_adr->isa_rawptr()) { 399 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 400 // Skip this node optimization if its address has TOP base. 401 return NodeSentinel; // caller will return null 402 } 403 404 // Avoid independent memory operations 405 Node* old_mem = mem; 406 407 // The code which unhooks non-raw memories from complete (macro-expanded) 408 // initializations was removed. After macro-expansion all stores caught 409 // by Initialize node became raw stores and there is no information 410 // which memory slices they modify. So it is unsafe to move any memory 411 // operation above these stores. Also in most cases hooked non-raw memories 412 // were already unhooked by using information from detect_ptr_independence() 413 // and find_previous_store(). 414 415 if (mem->is_MergeMem()) { 416 MergeMemNode* mmem = mem->as_MergeMem(); 417 const TypePtr *tp = t_adr->is_ptr(); 418 419 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 420 } 421 422 if (mem != old_mem) { 423 set_req_X(MemNode::Memory, mem, phase); 424 if (phase->type(mem) == Type::TOP) return NodeSentinel; 425 return this; 426 } 427 428 // let the subclass continue analyzing... 429 return nullptr; 430 } 431 432 // Helper function for proving some simple control dominations. 433 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 434 // Already assumes that 'dom' is available at 'sub', and that 'sub' 435 // is not a constant (dominated by the method's StartNode). 436 // Used by MemNode::find_previous_store to prove that the 437 // control input of a memory operation predates (dominates) 438 // an allocation it wants to look past. 439 // Returns 'DomResult::Dominate' if all control inputs of 'dom' 440 // dominate 'sub', 'DomResult::NotDominate' if not, 441 // and 'DomResult::EncounteredDeadCode' if we can't decide due to 442 // dead code, but at the end of IGVN, we know the definite result 443 // once the dead code is cleaned up. 444 Node::DomResult MemNode::maybe_all_controls_dominate(Node* dom, Node* sub) { 445 if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top()) { 446 return DomResult::EncounteredDeadCode; // Conservative answer for dead code 447 } 448 449 // Check 'dom'. Skip Proj and CatchProj nodes. 450 dom = dom->find_exact_control(dom); 451 if (dom == nullptr || dom->is_top()) { 452 return DomResult::EncounteredDeadCode; // Conservative answer for dead code 453 } 454 455 if (dom == sub) { 456 // For the case when, for example, 'sub' is Initialize and the original 457 // 'dom' is Proj node of the 'sub'. 458 return DomResult::NotDominate; 459 } 460 461 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) { 462 return DomResult::Dominate; 463 } 464 465 // 'dom' dominates 'sub' if its control edge and control edges 466 // of all its inputs dominate or equal to sub's control edge. 467 468 // Currently 'sub' is either Allocate, Initialize or Start nodes. 469 // Or Region for the check in LoadNode::Ideal(); 470 // 'sub' should have sub->in(0) != nullptr. 471 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 472 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 473 474 // Get control edge of 'sub'. 475 Node* orig_sub = sub; 476 sub = sub->find_exact_control(sub->in(0)); 477 if (sub == nullptr || sub->is_top()) { 478 return DomResult::EncounteredDeadCode; // Conservative answer for dead code 479 } 480 481 assert(sub->is_CFG(), "expecting control"); 482 483 if (sub == dom) { 484 return DomResult::Dominate; 485 } 486 487 if (sub->is_Start() || sub->is_Root()) { 488 return DomResult::NotDominate; 489 } 490 491 { 492 // Check all control edges of 'dom'. 493 494 ResourceMark rm; 495 Node_List nlist; 496 Unique_Node_List dom_list; 497 498 dom_list.push(dom); 499 bool only_dominating_controls = false; 500 501 for (uint next = 0; next < dom_list.size(); next++) { 502 Node* n = dom_list.at(next); 503 if (n == orig_sub) { 504 return DomResult::NotDominate; // One of dom's inputs dominated by sub. 505 } 506 if (!n->is_CFG() && n->pinned()) { 507 // Check only own control edge for pinned non-control nodes. 508 n = n->find_exact_control(n->in(0)); 509 if (n == nullptr || n->is_top()) { 510 return DomResult::EncounteredDeadCode; // Conservative answer for dead code 511 } 512 assert(n->is_CFG(), "expecting control"); 513 dom_list.push(n); 514 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 515 only_dominating_controls = true; 516 } else if (n->is_CFG()) { 517 DomResult dom_result = n->dominates(sub, nlist); 518 if (dom_result == DomResult::Dominate) { 519 only_dominating_controls = true; 520 } else { 521 return dom_result; 522 } 523 } else { 524 // First, own control edge. 525 Node* m = n->find_exact_control(n->in(0)); 526 if (m != nullptr) { 527 if (m->is_top()) { 528 return DomResult::EncounteredDeadCode; // Conservative answer for dead code 529 } 530 dom_list.push(m); 531 } 532 // Now, the rest of edges. 533 uint cnt = n->req(); 534 for (uint i = 1; i < cnt; i++) { 535 m = n->find_exact_control(n->in(i)); 536 if (m == nullptr || m->is_top()) { 537 continue; 538 } 539 dom_list.push(m); 540 } 541 } 542 } 543 return only_dominating_controls ? DomResult::Dominate : DomResult::NotDominate; 544 } 545 } 546 547 //---------------------detect_ptr_independence--------------------------------- 548 // Used by MemNode::find_previous_store to prove that two base 549 // pointers are never equal. 550 // The pointers are accompanied by their associated allocations, 551 // if any, which have been previously discovered by the caller. 552 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 553 Node* p2, AllocateNode* a2, 554 PhaseTransform* phase) { 555 // Attempt to prove that these two pointers cannot be aliased. 556 // They may both manifestly be allocations, and they should differ. 557 // Or, if they are not both allocations, they can be distinct constants. 558 // Otherwise, one is an allocation and the other a pre-existing value. 559 if (a1 == nullptr && a2 == nullptr) { // neither an allocation 560 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 561 } else if (a1 != nullptr && a2 != nullptr) { // both allocations 562 return (a1 != a2); 563 } else if (a1 != nullptr) { // one allocation a1 564 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 565 return all_controls_dominate(p2, a1); 566 } else { //(a2 != null) // one allocation a2 567 return all_controls_dominate(p1, a2); 568 } 569 return false; 570 } 571 572 573 // Find an arraycopy ac that produces the memory state represented by parameter mem. 574 // Return ac if 575 // (a) can_see_stored_value=true and ac must have set the value for this load or if 576 // (b) can_see_stored_value=false and ac could have set the value for this load or if 577 // (c) can_see_stored_value=false and ac cannot have set the value for this load. 578 // In case (c) change the parameter mem to the memory input of ac to skip it 579 // when searching stored value. 580 // Otherwise return null. 581 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const { 582 ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem); 583 if (ac != nullptr) { 584 Node* ld_addp = in(MemNode::Address); 585 Node* src = ac->in(ArrayCopyNode::Src); 586 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr(); 587 588 // This is a load from a cloned array. The corresponding arraycopy ac must 589 // have set the value for the load and we can return ac but only if the load 590 // is known to be within bounds. This is checked below. 591 if (ary_t != nullptr && ld_addp->is_AddP()) { 592 Node* ld_offs = ld_addp->in(AddPNode::Offset); 593 BasicType ary_elem = ary_t->elem()->array_element_basic_type(); 594 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem); 595 jlong elemsize = type2aelembytes(ary_elem); 596 597 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 598 const TypeInt* sizetype = ary_t->size(); 599 600 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) { 601 // The load is known to be within bounds. It receives its value from ac. 602 return ac; 603 } 604 // The load is known to be out-of-bounds. 605 } 606 // The load could be out-of-bounds. It must not be hoisted but must remain 607 // dependent on the runtime range check. This is achieved by returning null. 608 } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) { 609 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy(); 610 611 if (ac->is_arraycopy_validated() || 612 ac->is_copyof_validated() || 613 ac->is_copyofrange_validated()) { 614 Node* ld_addp = in(MemNode::Address); 615 if (ld_addp->is_AddP()) { 616 Node* ld_base = ld_addp->in(AddPNode::Address); 617 Node* ld_offs = ld_addp->in(AddPNode::Offset); 618 619 Node* dest = ac->in(ArrayCopyNode::Dest); 620 621 if (dest == ld_base) { 622 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 623 assert(!ld_offs_t->empty(), "dead reference should be checked already"); 624 // Take into account vector or unsafe access size 625 jlong ld_size_in_bytes = (jlong)memory_size(); 626 jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1; 627 offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM 628 if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) { 629 return ac; 630 } 631 if (!can_see_stored_value) { 632 mem = ac->in(TypeFunc::Memory); 633 return ac; 634 } 635 } 636 } 637 } 638 } 639 return nullptr; 640 } 641 642 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const { 643 if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore || 644 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) { 645 if (ld_alloc != nullptr) { 646 // Check if there is an array copy for a clone 647 Node* mb = mem->in(0); 648 ArrayCopyNode* ac = nullptr; 649 if (mb->in(0) != nullptr && mb->in(0)->is_Proj() && 650 mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) { 651 ac = mb->in(0)->in(0)->as_ArrayCopy(); 652 } else { 653 // Step over GC barrier when ReduceInitialCardMarks is disabled 654 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 655 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0)); 656 657 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) { 658 ac = control_proj_ac->in(0)->as_ArrayCopy(); 659 } 660 } 661 662 if (ac != nullptr && ac->is_clonebasic()) { 663 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest)); 664 if (alloc != nullptr && alloc == ld_alloc) { 665 return ac; 666 } 667 } 668 } 669 } 670 return nullptr; 671 } 672 673 // The logic for reordering loads and stores uses four steps: 674 // (a) Walk carefully past stores and initializations which we 675 // can prove are independent of this load. 676 // (b) Observe that the next memory state makes an exact match 677 // with self (load or store), and locate the relevant store. 678 // (c) Ensure that, if we were to wire self directly to the store, 679 // the optimizer would fold it up somehow. 680 // (d) Do the rewiring, and return, depending on some other part of 681 // the optimizer to fold up the load. 682 // This routine handles steps (a) and (b). Steps (c) and (d) are 683 // specific to loads and stores, so they are handled by the callers. 684 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 685 // 686 Node* MemNode::find_previous_store(PhaseValues* phase) { 687 Node* ctrl = in(MemNode::Control); 688 Node* adr = in(MemNode::Address); 689 intptr_t offset = 0; 690 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 691 AllocateNode* alloc = AllocateNode::Ideal_allocation(base); 692 693 if (offset == Type::OffsetBot) 694 return nullptr; // cannot unalias unless there are precise offsets 695 696 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr); 697 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 698 699 intptr_t size_in_bytes = memory_size(); 700 701 Node* mem = in(MemNode::Memory); // start searching here... 702 703 int cnt = 50; // Cycle limiter 704 for (;;) { // While we can dance past unrelated stores... 705 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 706 707 Node* prev = mem; 708 if (mem->is_Store()) { 709 Node* st_adr = mem->in(MemNode::Address); 710 intptr_t st_offset = 0; 711 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 712 if (st_base == nullptr) 713 break; // inscrutable pointer 714 715 // For raw accesses it's not enough to prove that constant offsets don't intersect. 716 // We need the bases to be the equal in order for the offset check to make sense. 717 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) { 718 break; 719 } 720 721 if (st_offset != offset && st_offset != Type::OffsetBot) { 722 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize); 723 assert(mem->as_Store()->memory_size() <= MAX_STORE, ""); 724 if (st_offset >= offset + size_in_bytes || 725 st_offset <= offset - MAX_STORE || 726 st_offset <= offset - mem->as_Store()->memory_size()) { 727 // Success: The offsets are provably independent. 728 // (You may ask, why not just test st_offset != offset and be done? 729 // The answer is that stores of different sizes can co-exist 730 // in the same sequence of RawMem effects. We sometimes initialize 731 // a whole 'tile' of array elements with a single jint or jlong.) 732 mem = mem->in(MemNode::Memory); 733 continue; // (a) advance through independent store memory 734 } 735 } 736 if (st_base != base && 737 detect_ptr_independence(base, alloc, 738 st_base, 739 AllocateNode::Ideal_allocation(st_base), 740 phase)) { 741 // Success: The bases are provably independent. 742 mem = mem->in(MemNode::Memory); 743 continue; // (a) advance through independent store memory 744 } 745 746 // (b) At this point, if the bases or offsets do not agree, we lose, 747 // since we have not managed to prove 'this' and 'mem' independent. 748 if (st_base == base && st_offset == offset) { 749 return mem; // let caller handle steps (c), (d) 750 } 751 752 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 753 InitializeNode* st_init = mem->in(0)->as_Initialize(); 754 AllocateNode* st_alloc = st_init->allocation(); 755 if (st_alloc == nullptr) { 756 break; // something degenerated 757 } 758 bool known_identical = false; 759 bool known_independent = false; 760 if (alloc == st_alloc) { 761 known_identical = true; 762 } else if (alloc != nullptr) { 763 known_independent = true; 764 } else if (all_controls_dominate(this, st_alloc)) { 765 known_independent = true; 766 } 767 768 if (known_independent) { 769 // The bases are provably independent: Either they are 770 // manifestly distinct allocations, or else the control 771 // of this load dominates the store's allocation. 772 int alias_idx = phase->C->get_alias_index(adr_type()); 773 if (alias_idx == Compile::AliasIdxRaw) { 774 mem = st_alloc->in(TypeFunc::Memory); 775 } else { 776 mem = st_init->memory(alias_idx); 777 } 778 continue; // (a) advance through independent store memory 779 } 780 781 // (b) at this point, if we are not looking at a store initializing 782 // the same allocation we are loading from, we lose. 783 if (known_identical) { 784 // From caller, can_see_stored_value will consult find_captured_store. 785 return mem; // let caller handle steps (c), (d) 786 } 787 788 } else if (find_previous_arraycopy(phase, alloc, mem, false) != nullptr) { 789 if (prev != mem) { 790 // Found an arraycopy but it doesn't affect that load 791 continue; 792 } 793 // Found an arraycopy that may affect that load 794 return mem; 795 } else if (addr_t != nullptr && addr_t->is_known_instance_field()) { 796 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 797 if (mem->is_Proj() && mem->in(0)->is_Call()) { 798 // ArrayCopyNodes processed here as well. 799 CallNode *call = mem->in(0)->as_Call(); 800 if (!call->may_modify(addr_t, phase)) { 801 mem = call->in(TypeFunc::Memory); 802 continue; // (a) advance through independent call memory 803 } 804 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 805 ArrayCopyNode* ac = nullptr; 806 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) { 807 break; 808 } 809 mem = mem->in(0)->in(TypeFunc::Memory); 810 continue; // (a) advance through independent MemBar memory 811 } else if (mem->is_ClearArray()) { 812 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 813 // (the call updated 'mem' value) 814 continue; // (a) advance through independent allocation memory 815 } else { 816 // Can not bypass initialization of the instance 817 // we are looking for. 818 return mem; 819 } 820 } else if (mem->is_MergeMem()) { 821 int alias_idx = phase->C->get_alias_index(adr_type()); 822 mem = mem->as_MergeMem()->memory_at(alias_idx); 823 continue; // (a) advance through independent MergeMem memory 824 } 825 } 826 827 // Unless there is an explicit 'continue', we must bail out here, 828 // because 'mem' is an inscrutable memory state (e.g., a call). 829 break; 830 } 831 832 return nullptr; // bail out 833 } 834 835 //----------------------calculate_adr_type------------------------------------- 836 // Helper function. Notices when the given type of address hits top or bottom. 837 // Also, asserts a cross-check of the type against the expected address type. 838 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 839 if (t == Type::TOP) return nullptr; // does not touch memory any more? 840 #ifdef ASSERT 841 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = nullptr; 842 #endif 843 const TypePtr* tp = t->isa_ptr(); 844 if (tp == nullptr) { 845 assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 846 return TypePtr::BOTTOM; // touches lots of memory 847 } else { 848 #ifdef ASSERT 849 // %%%% [phh] We don't check the alias index if cross_check is 850 // TypeRawPtr::BOTTOM. Needs to be investigated. 851 if (cross_check != nullptr && 852 cross_check != TypePtr::BOTTOM && 853 cross_check != TypeRawPtr::BOTTOM) { 854 // Recheck the alias index, to see if it has changed (due to a bug). 855 Compile* C = Compile::current(); 856 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 857 "must stay in the original alias category"); 858 // The type of the address must be contained in the adr_type, 859 // disregarding "null"-ness. 860 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 861 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 862 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 863 "real address must not escape from expected memory type"); 864 } 865 #endif 866 return tp; 867 } 868 } 869 870 uint8_t MemNode::barrier_data(const Node* n) { 871 if (n->is_LoadStore()) { 872 return n->as_LoadStore()->barrier_data(); 873 } else if (n->is_Mem()) { 874 return n->as_Mem()->barrier_data(); 875 } 876 return 0; 877 } 878 879 //============================================================================= 880 // Should LoadNode::Ideal() attempt to remove control edges? 881 bool LoadNode::can_remove_control() const { 882 return !has_pinned_control_dependency(); 883 } 884 uint LoadNode::size_of() const { return sizeof(*this); } 885 bool LoadNode::cmp(const Node &n) const { 886 LoadNode& load = (LoadNode &)n; 887 return Type::equals(_type, load._type) && 888 _control_dependency == load._control_dependency && 889 _mo == load._mo; 890 } 891 const Type *LoadNode::bottom_type() const { return _type; } 892 uint LoadNode::ideal_reg() const { 893 return _type->ideal_reg(); 894 } 895 896 #ifndef PRODUCT 897 void LoadNode::dump_spec(outputStream *st) const { 898 MemNode::dump_spec(st); 899 if( !Verbose && !WizardMode ) { 900 // standard dump does this in Verbose and WizardMode 901 st->print(" #"); _type->dump_on(st); 902 } 903 if (!depends_only_on_test()) { 904 st->print(" (does not depend only on test, "); 905 if (control_dependency() == UnknownControl) { 906 st->print("unknown control"); 907 } else if (control_dependency() == Pinned) { 908 st->print("pinned"); 909 } else if (adr_type() == TypeRawPtr::BOTTOM) { 910 st->print("raw access"); 911 } else { 912 st->print("unknown reason"); 913 } 914 st->print(")"); 915 } 916 } 917 #endif 918 919 #ifdef ASSERT 920 //----------------------------is_immutable_value------------------------------- 921 // Helper function to allow a raw load without control edge for some cases 922 bool LoadNode::is_immutable_value(Node* adr) { 923 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 924 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) { 925 926 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1); 927 int offsets[] = { 928 in_bytes(JavaThread::osthread_offset()), 929 in_bytes(JavaThread::threadObj_offset()), 930 in_bytes(JavaThread::vthread_offset()), 931 in_bytes(JavaThread::scopedValueCache_offset()), 932 }; 933 934 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) { 935 if (offset == offsets[i]) { 936 return true; 937 } 938 } 939 } 940 941 return false; 942 } 943 #endif 944 945 //----------------------------LoadNode::make----------------------------------- 946 // Polymorphic factory method: 947 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo, 948 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) { 949 Compile* C = gvn.C; 950 951 // sanity check the alias category against the created node type 952 assert(!(adr_type->isa_oopptr() && 953 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 954 "use LoadKlassNode instead"); 955 assert(!(adr_type->isa_aryptr() && 956 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 957 "use LoadRangeNode instead"); 958 // Check control edge of raw loads 959 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 960 // oop will be recorded in oop map if load crosses safepoint 961 rt->isa_oopptr() || is_immutable_value(adr), 962 "raw memory operations should have control edge"); 963 LoadNode* load = nullptr; 964 switch (bt) { 965 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 966 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 967 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 968 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 969 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 970 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break; 971 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 972 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break; 973 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break; 974 case T_OBJECT: 975 case T_NARROWOOP: 976 #ifdef _LP64 977 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 978 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency); 979 } else 980 #endif 981 { 982 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 983 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 984 } 985 break; 986 default: 987 ShouldNotReachHere(); 988 break; 989 } 990 assert(load != nullptr, "LoadNode should have been created"); 991 if (unaligned) { 992 load->set_unaligned_access(); 993 } 994 if (mismatched) { 995 load->set_mismatched_access(); 996 } 997 if (unsafe) { 998 load->set_unsafe_access(); 999 } 1000 load->set_barrier_data(barrier_data); 1001 if (load->Opcode() == Op_LoadN) { 1002 Node* ld = gvn.transform(load); 1003 return new DecodeNNode(ld, ld->bottom_type()->make_ptr()); 1004 } 1005 1006 return load; 1007 } 1008 1009 //------------------------------hash------------------------------------------- 1010 uint LoadNode::hash() const { 1011 // unroll addition of interesting fields 1012 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 1013 } 1014 1015 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 1016 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) { 1017 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile(); 1018 bool is_stable_ary = FoldStableValues && 1019 (tp != nullptr) && (tp->isa_aryptr() != nullptr) && 1020 tp->isa_aryptr()->is_stable(); 1021 1022 return (eliminate_boxing && non_volatile) || is_stable_ary; 1023 } 1024 1025 return false; 1026 } 1027 1028 LoadNode* LoadNode::pin_array_access_node() const { 1029 const TypePtr* adr_type = this->adr_type(); 1030 if (adr_type != nullptr && adr_type->isa_aryptr()) { 1031 return clone_pinned(); 1032 } 1033 return nullptr; 1034 } 1035 1036 // Is the value loaded previously stored by an arraycopy? If so return 1037 // a load node that reads from the source array so we may be able to 1038 // optimize out the ArrayCopy node later. 1039 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const { 1040 Node* ld_adr = in(MemNode::Address); 1041 intptr_t ld_off = 0; 1042 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 1043 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 1044 if (ac != nullptr) { 1045 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 1046 1047 Node* mem = ac->in(TypeFunc::Memory); 1048 Node* ctl = ac->in(0); 1049 Node* src = ac->in(ArrayCopyNode::Src); 1050 1051 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) { 1052 return nullptr; 1053 } 1054 1055 // load depends on the tests that validate the arraycopy 1056 LoadNode* ld = clone_pinned(); 1057 Node* addp = in(MemNode::Address)->clone(); 1058 if (ac->as_ArrayCopy()->is_clonebasic()) { 1059 assert(ld_alloc != nullptr, "need an alloc"); 1060 assert(addp->is_AddP(), "address must be addp"); 1061 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1062 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern"); 1063 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern"); 1064 addp->set_req(AddPNode::Base, src); 1065 addp->set_req(AddPNode::Address, src); 1066 } else { 1067 assert(ac->as_ArrayCopy()->is_arraycopy_validated() || 1068 ac->as_ArrayCopy()->is_copyof_validated() || 1069 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases"); 1070 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 1071 addp->set_req(AddPNode::Base, src); 1072 addp->set_req(AddPNode::Address, src); 1073 1074 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 1075 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type(); 1076 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT; 1077 1078 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 1079 uint shift = exact_log2(type2aelembytes(ary_elem)); 1080 1081 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 1082 #ifdef _LP64 1083 diff = phase->transform(new ConvI2LNode(diff)); 1084 #endif 1085 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 1086 1087 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 1088 addp->set_req(AddPNode::Offset, offset); 1089 } 1090 addp = phase->transform(addp); 1091 #ifdef ASSERT 1092 const TypePtr* adr_type = phase->type(addp)->is_ptr(); 1093 ld->_adr_type = adr_type; 1094 #endif 1095 ld->set_req(MemNode::Address, addp); 1096 ld->set_req(0, ctl); 1097 ld->set_req(MemNode::Memory, mem); 1098 return ld; 1099 } 1100 return nullptr; 1101 } 1102 1103 1104 //---------------------------can_see_stored_value------------------------------ 1105 // This routine exists to make sure this set of tests is done the same 1106 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 1107 // will change the graph shape in a way which makes memory alive twice at the 1108 // same time (uses the Oracle model of aliasing), then some 1109 // LoadXNode::Identity will fold things back to the equivalence-class model 1110 // of aliasing. 1111 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const { 1112 Node* ld_adr = in(MemNode::Address); 1113 intptr_t ld_off = 0; 1114 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off); 1115 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base); 1116 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 1117 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr; 1118 // This is more general than load from boxing objects. 1119 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 1120 uint alias_idx = atp->index(); 1121 Node* result = nullptr; 1122 Node* current = st; 1123 // Skip through chains of MemBarNodes checking the MergeMems for 1124 // new states for the slice of this load. Stop once any other 1125 // kind of node is encountered. Loads from final memory can skip 1126 // through any kind of MemBar but normal loads shouldn't skip 1127 // through MemBarAcquire since the could allow them to move out of 1128 // a synchronized region. It is not safe to step over MemBarCPUOrder, 1129 // because alias info above them may be inaccurate (e.g., due to 1130 // mixed/mismatched unsafe accesses). 1131 bool is_final_mem = !atp->is_rewritable(); 1132 while (current->is_Proj()) { 1133 int opc = current->in(0)->Opcode(); 1134 if ((is_final_mem && (opc == Op_MemBarAcquire || 1135 opc == Op_MemBarAcquireLock || 1136 opc == Op_LoadFence)) || 1137 opc == Op_MemBarRelease || 1138 opc == Op_StoreFence || 1139 opc == Op_MemBarReleaseLock || 1140 opc == Op_MemBarStoreStore || 1141 opc == Op_StoreStoreFence) { 1142 Node* mem = current->in(0)->in(TypeFunc::Memory); 1143 if (mem->is_MergeMem()) { 1144 MergeMemNode* merge = mem->as_MergeMem(); 1145 Node* new_st = merge->memory_at(alias_idx); 1146 if (new_st == merge->base_memory()) { 1147 // Keep searching 1148 current = new_st; 1149 continue; 1150 } 1151 // Save the new memory state for the slice and fall through 1152 // to exit. 1153 result = new_st; 1154 } 1155 } 1156 break; 1157 } 1158 if (result != nullptr) { 1159 st = result; 1160 } 1161 } 1162 1163 // Loop around twice in the case Load -> Initialize -> Store. 1164 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 1165 for (int trip = 0; trip <= 1; trip++) { 1166 1167 if (st->is_Store()) { 1168 Node* st_adr = st->in(MemNode::Address); 1169 if (st_adr != ld_adr) { 1170 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers). 1171 intptr_t st_off = 0; 1172 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off); 1173 if (ld_base == nullptr) return nullptr; 1174 if (st_base == nullptr) return nullptr; 1175 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr; 1176 if (ld_off != st_off) return nullptr; 1177 if (ld_off == Type::OffsetBot) return nullptr; 1178 // Same base, same offset. 1179 // Possible improvement for arrays: check index value instead of absolute offset. 1180 1181 // At this point we have proven something like this setup: 1182 // B = << base >> 1183 // L = LoadQ(AddP(Check/CastPP(B), #Off)) 1184 // S = StoreQ(AddP( B , #Off), V) 1185 // (Actually, we haven't yet proven the Q's are the same.) 1186 // In other words, we are loading from a casted version of 1187 // the same pointer-and-offset that we stored to. 1188 // Casted version may carry a dependency and it is respected. 1189 // Thus, we are able to replace L by V. 1190 } 1191 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1192 if (store_Opcode() != st->Opcode()) { 1193 return nullptr; 1194 } 1195 // LoadVector/StoreVector needs additional check to ensure the types match. 1196 if (st->is_StoreVector()) { 1197 const TypeVect* in_vt = st->as_StoreVector()->vect_type(); 1198 const TypeVect* out_vt = as_LoadVector()->vect_type(); 1199 if (in_vt != out_vt) { 1200 return nullptr; 1201 } 1202 } 1203 return st->in(MemNode::ValueIn); 1204 } 1205 1206 // A load from a freshly-created object always returns zero. 1207 // (This can happen after LoadNode::Ideal resets the load's memory input 1208 // to find_captured_store, which returned InitializeNode::zero_memory.) 1209 if (st->is_Proj() && st->in(0)->is_Allocate() && 1210 (st->in(0) == ld_alloc) && 1211 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1212 // return a zero value for the load's basic type 1213 // (This is one of the few places where a generic PhaseTransform 1214 // can create new nodes. Think of it as lazily manifesting 1215 // virtually pre-existing constants.) 1216 if (value_basic_type() != T_VOID) { 1217 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) { 1218 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an 1219 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done 1220 // by the ArrayCopyNode. 1221 return phase->zerocon(value_basic_type()); 1222 } 1223 } else { 1224 // TODO: materialize all-zero vector constant 1225 assert(!isa_Load() || as_Load()->type()->isa_vect(), ""); 1226 } 1227 } 1228 1229 // A load from an initialization barrier can match a captured store. 1230 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1231 InitializeNode* init = st->in(0)->as_Initialize(); 1232 AllocateNode* alloc = init->allocation(); 1233 if ((alloc != nullptr) && (alloc == ld_alloc)) { 1234 // examine a captured store value 1235 st = init->find_captured_store(ld_off, memory_size(), phase); 1236 if (st != nullptr) { 1237 continue; // take one more trip around 1238 } 1239 } 1240 } 1241 1242 // Load boxed value from result of valueOf() call is input parameter. 1243 if (this->is_Load() && ld_adr->is_AddP() && 1244 (tp != nullptr) && tp->is_ptr_to_boxed_value()) { 1245 intptr_t ignore = 0; 1246 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1247 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1248 base = bs->step_over_gc_barrier(base); 1249 if (base != nullptr && base->is_Proj() && 1250 base->as_Proj()->_con == TypeFunc::Parms && 1251 base->in(0)->is_CallStaticJava() && 1252 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1253 return base->in(0)->in(TypeFunc::Parms); 1254 } 1255 } 1256 1257 break; 1258 } 1259 1260 return nullptr; 1261 } 1262 1263 //----------------------is_instance_field_load_with_local_phi------------------ 1264 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1265 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1266 in(Address)->is_AddP() ) { 1267 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1268 // Only instances and boxed values. 1269 if( t_oop != nullptr && 1270 (t_oop->is_ptr_to_boxed_value() || 1271 t_oop->is_known_instance_field()) && 1272 t_oop->offset() != Type::OffsetBot && 1273 t_oop->offset() != Type::OffsetTop) { 1274 return true; 1275 } 1276 } 1277 return false; 1278 } 1279 1280 //------------------------------Identity--------------------------------------- 1281 // Loads are identity if previous store is to same address 1282 Node* LoadNode::Identity(PhaseGVN* phase) { 1283 // If the previous store-maker is the right kind of Store, and the store is 1284 // to the same address, then we are equal to the value stored. 1285 Node* mem = in(Memory); 1286 Node* value = can_see_stored_value(mem, phase); 1287 if( value ) { 1288 // byte, short & char stores truncate naturally. 1289 // A load has to load the truncated value which requires 1290 // some sort of masking operation and that requires an 1291 // Ideal call instead of an Identity call. 1292 if (memory_size() < BytesPerInt) { 1293 // If the input to the store does not fit with the load's result type, 1294 // it must be truncated via an Ideal call. 1295 if (!phase->type(value)->higher_equal(phase->type(this))) 1296 return this; 1297 } 1298 // (This works even when value is a Con, but LoadNode::Value 1299 // usually runs first, producing the singleton type of the Con.) 1300 if (!has_pinned_control_dependency() || value->is_Con()) { 1301 return value; 1302 } else { 1303 return this; 1304 } 1305 } 1306 1307 if (has_pinned_control_dependency()) { 1308 return this; 1309 } 1310 // Search for an existing data phi which was generated before for the same 1311 // instance's field to avoid infinite generation of phis in a loop. 1312 Node *region = mem->in(0); 1313 if (is_instance_field_load_with_local_phi(region)) { 1314 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1315 int this_index = phase->C->get_alias_index(addr_t); 1316 int this_offset = addr_t->offset(); 1317 int this_iid = addr_t->instance_id(); 1318 if (!addr_t->is_known_instance() && 1319 addr_t->is_ptr_to_boxed_value()) { 1320 // Use _idx of address base (could be Phi node) for boxed values. 1321 intptr_t ignore = 0; 1322 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1323 if (base == nullptr) { 1324 return this; 1325 } 1326 this_iid = base->_idx; 1327 } 1328 const Type* this_type = bottom_type(); 1329 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1330 Node* phi = region->fast_out(i); 1331 if (phi->is_Phi() && phi != mem && 1332 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1333 return phi; 1334 } 1335 } 1336 } 1337 1338 return this; 1339 } 1340 1341 // Construct an equivalent unsigned load. 1342 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) { 1343 BasicType bt = T_ILLEGAL; 1344 const Type* rt = nullptr; 1345 switch (Opcode()) { 1346 case Op_LoadUB: return this; 1347 case Op_LoadUS: return this; 1348 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break; 1349 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break; 1350 default: 1351 assert(false, "no unsigned variant: %s", Name()); 1352 return nullptr; 1353 } 1354 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1355 raw_adr_type(), rt, bt, _mo, _control_dependency, 1356 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access()); 1357 } 1358 1359 // Construct an equivalent signed load. 1360 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) { 1361 BasicType bt = T_ILLEGAL; 1362 const Type* rt = nullptr; 1363 switch (Opcode()) { 1364 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break; 1365 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break; 1366 case Op_LoadB: // fall through 1367 case Op_LoadS: // fall through 1368 case Op_LoadI: // fall through 1369 case Op_LoadL: return this; 1370 default: 1371 assert(false, "no signed variant: %s", Name()); 1372 return nullptr; 1373 } 1374 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1375 raw_adr_type(), rt, bt, _mo, _control_dependency, 1376 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access()); 1377 } 1378 1379 bool LoadNode::has_reinterpret_variant(const Type* rt) { 1380 BasicType bt = rt->basic_type(); 1381 switch (Opcode()) { 1382 case Op_LoadI: return (bt == T_FLOAT); 1383 case Op_LoadL: return (bt == T_DOUBLE); 1384 case Op_LoadF: return (bt == T_INT); 1385 case Op_LoadD: return (bt == T_LONG); 1386 1387 default: return false; 1388 } 1389 } 1390 1391 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) { 1392 BasicType bt = rt->basic_type(); 1393 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt)); 1394 bool is_mismatched = is_mismatched_access(); 1395 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr(); 1396 if (raw_type == nullptr) { 1397 is_mismatched = true; // conservatively match all non-raw accesses as mismatched 1398 } 1399 const int op = Opcode(); 1400 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) || 1401 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access()); 1402 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1403 raw_adr_type(), rt, bt, _mo, _control_dependency, 1404 require_atomic_access, is_unaligned_access(), is_mismatched); 1405 } 1406 1407 bool StoreNode::has_reinterpret_variant(const Type* vt) { 1408 BasicType bt = vt->basic_type(); 1409 switch (Opcode()) { 1410 case Op_StoreI: return (bt == T_FLOAT); 1411 case Op_StoreL: return (bt == T_DOUBLE); 1412 case Op_StoreF: return (bt == T_INT); 1413 case Op_StoreD: return (bt == T_LONG); 1414 1415 default: return false; 1416 } 1417 } 1418 1419 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) { 1420 BasicType bt = vt->basic_type(); 1421 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt)); 1422 const int op = Opcode(); 1423 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) || 1424 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access()); 1425 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1426 raw_adr_type(), val, bt, _mo, require_atomic_access); 1427 1428 bool is_mismatched = is_mismatched_access(); 1429 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr(); 1430 if (raw_type == nullptr) { 1431 is_mismatched = true; // conservatively match all non-raw accesses as mismatched 1432 } 1433 if (is_mismatched) { 1434 st->set_mismatched_access(); 1435 } 1436 return st; 1437 } 1438 1439 // We're loading from an object which has autobox behaviour. 1440 // If this object is result of a valueOf call we'll have a phi 1441 // merging a newly allocated object and a load from the cache. 1442 // We want to replace this load with the original incoming 1443 // argument to the valueOf call. 1444 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) { 1445 assert(igvn->C->eliminate_boxing(), "sanity"); 1446 intptr_t ignore = 0; 1447 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore); 1448 if ((base == nullptr) || base->is_Phi()) { 1449 // Push the loads from the phi that comes from valueOf up 1450 // through it to allow elimination of the loads and the recovery 1451 // of the original value. It is done in split_through_phi(). 1452 return nullptr; 1453 } else if (base->is_Load() || 1454 (base->is_DecodeN() && base->in(1)->is_Load())) { 1455 // Eliminate the load of boxed value for integer types from the cache 1456 // array by deriving the value from the index into the array. 1457 // Capture the offset of the load and then reverse the computation. 1458 1459 // Get LoadN node which loads a boxing object from 'cache' array. 1460 if (base->is_DecodeN()) { 1461 base = base->in(1); 1462 } 1463 if (!base->in(Address)->is_AddP()) { 1464 return nullptr; // Complex address 1465 } 1466 AddPNode* address = base->in(Address)->as_AddP(); 1467 Node* cache_base = address->in(AddPNode::Base); 1468 if ((cache_base != nullptr) && cache_base->is_DecodeN()) { 1469 // Get ConP node which is static 'cache' field. 1470 cache_base = cache_base->in(1); 1471 } 1472 if ((cache_base != nullptr) && cache_base->is_Con()) { 1473 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1474 if ((base_type != nullptr) && base_type->is_autobox_cache()) { 1475 Node* elements[4]; 1476 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1477 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1478 if (count > 0 && elements[0]->is_Con() && 1479 (count == 1 || 1480 (count == 2 && elements[1]->Opcode() == Op_LShiftX && 1481 elements[1]->in(2) == igvn->intcon(shift)))) { 1482 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1483 // Fetch the box object cache[0] at the base of the array and get its value 1484 ciInstance* box = array->obj_at(0)->as_instance(); 1485 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1486 assert(ik->is_box_klass(), "sanity"); 1487 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1488 if (ik->nof_nonstatic_fields() == 1) { 1489 // This should be true nonstatic_field_at requires calling 1490 // nof_nonstatic_fields so check it anyway 1491 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1492 BasicType bt = c.basic_type(); 1493 // Only integer types have boxing cache. 1494 assert(bt == T_BOOLEAN || bt == T_CHAR || 1495 bt == T_BYTE || bt == T_SHORT || 1496 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt)); 1497 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1498 if (cache_low != (int)cache_low) { 1499 return nullptr; // should not happen since cache is array indexed by value 1500 } 1501 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1502 if (offset != (int)offset) { 1503 return nullptr; // should not happen since cache is array indexed by value 1504 } 1505 // Add up all the offsets making of the address of the load 1506 Node* result = elements[0]; 1507 for (int i = 1; i < count; i++) { 1508 result = igvn->transform(new AddXNode(result, elements[i])); 1509 } 1510 // Remove the constant offset from the address and then 1511 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset))); 1512 // remove the scaling of the offset to recover the original index. 1513 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) { 1514 // Peel the shift off directly but wrap it in a dummy node 1515 // since Ideal can't return existing nodes 1516 igvn->_worklist.push(result); // remove dead node later 1517 result = new RShiftXNode(result->in(1), igvn->intcon(0)); 1518 } else if (result->is_Add() && result->in(2)->is_Con() && 1519 result->in(1)->Opcode() == Op_LShiftX && 1520 result->in(1)->in(2) == igvn->intcon(shift)) { 1521 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1522 // but for boxing cache access we know that X<<Z will not overflow 1523 // (there is range check) so we do this optimizatrion by hand here. 1524 igvn->_worklist.push(result); // remove dead node later 1525 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift)); 1526 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con)); 1527 } else { 1528 result = new RShiftXNode(result, igvn->intcon(shift)); 1529 } 1530 #ifdef _LP64 1531 if (bt != T_LONG) { 1532 result = new ConvL2INode(igvn->transform(result)); 1533 } 1534 #else 1535 if (bt == T_LONG) { 1536 result = new ConvI2LNode(igvn->transform(result)); 1537 } 1538 #endif 1539 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1540 // Need to preserve unboxing load type if it is unsigned. 1541 switch(this->Opcode()) { 1542 case Op_LoadUB: 1543 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF)); 1544 break; 1545 case Op_LoadUS: 1546 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF)); 1547 break; 1548 } 1549 return result; 1550 } 1551 } 1552 } 1553 } 1554 } 1555 return nullptr; 1556 } 1557 1558 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1559 Node* region = phi->in(0); 1560 if (region == nullptr) { 1561 return false; // Wait stable graph 1562 } 1563 uint cnt = phi->req(); 1564 for (uint i = 1; i < cnt; i++) { 1565 Node* rc = region->in(i); 1566 if (rc == nullptr || phase->type(rc) == Type::TOP) 1567 return false; // Wait stable graph 1568 Node* in = phi->in(i); 1569 if (in == nullptr || phase->type(in) == Type::TOP) 1570 return false; // Wait stable graph 1571 } 1572 return true; 1573 } 1574 1575 //------------------------------split_through_phi------------------------------ 1576 // Check whether a call to 'split_through_phi' would split this load through the 1577 // Phi *base*. This method is essentially a copy of the validations performed 1578 // by 'split_through_phi'. The first use of this method was in EA code as part 1579 // of simplification of allocation merges. 1580 // Some differences from original method (split_through_phi): 1581 // - If base->is_CastPP(): base = base->in(1) 1582 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) { 1583 Node* mem = in(Memory); 1584 Node* address = in(Address); 1585 intptr_t ignore = 0; 1586 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1587 1588 if (base->is_CastPP()) { 1589 base = base->in(1); 1590 } 1591 1592 if (req() > 3 || base == nullptr || !base->is_Phi()) { 1593 return false; 1594 } 1595 1596 if (!mem->is_Phi()) { 1597 if (!MemNode::all_controls_dominate(mem, base->in(0))) { 1598 return false; 1599 } 1600 } else if (base->in(0) != mem->in(0)) { 1601 if (!MemNode::all_controls_dominate(mem, base->in(0))) { 1602 return false; 1603 } 1604 } 1605 1606 return true; 1607 } 1608 1609 //------------------------------split_through_phi------------------------------ 1610 // Split instance or boxed field load through Phi. 1611 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) { 1612 if (req() > 3) { 1613 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs"); 1614 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account 1615 return nullptr; 1616 } 1617 Node* mem = in(Memory); 1618 Node* address = in(Address); 1619 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1620 1621 assert((t_oop != nullptr) && 1622 (ignore_missing_instance_id || 1623 t_oop->is_known_instance_field() || 1624 t_oop->is_ptr_to_boxed_value()), "invalid conditions"); 1625 1626 Compile* C = phase->C; 1627 intptr_t ignore = 0; 1628 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1629 bool base_is_phi = (base != nullptr) && base->is_Phi(); 1630 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1631 (base != nullptr) && (base == address->in(AddPNode::Base)) && 1632 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1633 1634 if (!((mem->is_Phi() || base_is_phi) && 1635 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) { 1636 return nullptr; // Neither memory or base are Phi 1637 } 1638 1639 if (mem->is_Phi()) { 1640 if (!stable_phi(mem->as_Phi(), phase)) { 1641 return nullptr; // Wait stable graph 1642 } 1643 uint cnt = mem->req(); 1644 // Check for loop invariant memory. 1645 if (cnt == 3) { 1646 for (uint i = 1; i < cnt; i++) { 1647 Node* in = mem->in(i); 1648 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1649 if (m == mem) { 1650 if (i == 1) { 1651 // if the first edge was a loop, check second edge too. 1652 // If both are replaceable - we are in an infinite loop 1653 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase); 1654 if (n == mem) { 1655 break; 1656 } 1657 } 1658 set_req(Memory, mem->in(cnt - i)); 1659 return this; // made change 1660 } 1661 } 1662 } 1663 } 1664 if (base_is_phi) { 1665 if (!stable_phi(base->as_Phi(), phase)) { 1666 return nullptr; // Wait stable graph 1667 } 1668 uint cnt = base->req(); 1669 // Check for loop invariant memory. 1670 if (cnt == 3) { 1671 for (uint i = 1; i < cnt; i++) { 1672 if (base->in(i) == base) { 1673 return nullptr; // Wait stable graph 1674 } 1675 } 1676 } 1677 } 1678 1679 // Split through Phi (see original code in loopopts.cpp). 1680 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type"); 1681 1682 // Do nothing here if Identity will find a value 1683 // (to avoid infinite chain of value phis generation). 1684 if (this != Identity(phase)) { 1685 return nullptr; 1686 } 1687 1688 // Select Region to split through. 1689 Node* region; 1690 DomResult dom_result = DomResult::Dominate; 1691 if (!base_is_phi) { 1692 assert(mem->is_Phi(), "sanity"); 1693 region = mem->in(0); 1694 // Skip if the region dominates some control edge of the address. 1695 // We will check `dom_result` later. 1696 dom_result = MemNode::maybe_all_controls_dominate(address, region); 1697 } else if (!mem->is_Phi()) { 1698 assert(base_is_phi, "sanity"); 1699 region = base->in(0); 1700 // Skip if the region dominates some control edge of the memory. 1701 // We will check `dom_result` later. 1702 dom_result = MemNode::maybe_all_controls_dominate(mem, region); 1703 } else if (base->in(0) != mem->in(0)) { 1704 assert(base_is_phi && mem->is_Phi(), "sanity"); 1705 dom_result = MemNode::maybe_all_controls_dominate(mem, base->in(0)); 1706 if (dom_result == DomResult::Dominate) { 1707 region = base->in(0); 1708 } else { 1709 dom_result = MemNode::maybe_all_controls_dominate(address, mem->in(0)); 1710 if (dom_result == DomResult::Dominate) { 1711 region = mem->in(0); 1712 } 1713 // Otherwise we encountered a complex graph. 1714 } 1715 } else { 1716 assert(base->in(0) == mem->in(0), "sanity"); 1717 region = mem->in(0); 1718 } 1719 1720 PhaseIterGVN* igvn = phase->is_IterGVN(); 1721 if (dom_result != DomResult::Dominate) { 1722 if (dom_result == DomResult::EncounteredDeadCode) { 1723 // There is some dead code which eventually will be removed in IGVN. 1724 // Once this is the case, we get an unambiguous dominance result. 1725 // Push the node to the worklist again until the dead code is removed. 1726 igvn->_worklist.push(this); 1727 } 1728 return nullptr; 1729 } 1730 1731 Node* phi = nullptr; 1732 const Type* this_type = this->bottom_type(); 1733 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) { 1734 int this_index = C->get_alias_index(t_oop); 1735 int this_offset = t_oop->offset(); 1736 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx; 1737 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset); 1738 } else if (ignore_missing_instance_id) { 1739 phi = new PhiNode(region, this_type, nullptr, mem->_idx); 1740 } else { 1741 return nullptr; 1742 } 1743 1744 for (uint i = 1; i < region->req(); i++) { 1745 Node* x; 1746 Node* the_clone = nullptr; 1747 Node* in = region->in(i); 1748 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl && 1749 in != nullptr && in->is_OuterStripMinedLoop()) { 1750 // No node should go in the outer strip mined loop 1751 in = in->in(LoopNode::EntryControl); 1752 } 1753 if (in == nullptr || in == C->top()) { 1754 x = C->top(); // Dead path? Use a dead data op 1755 } else { 1756 x = this->clone(); // Else clone up the data op 1757 the_clone = x; // Remember for possible deletion. 1758 // Alter data node to use pre-phi inputs 1759 if (this->in(0) == region) { 1760 x->set_req(0, in); 1761 } else { 1762 x->set_req(0, nullptr); 1763 } 1764 if (mem->is_Phi() && (mem->in(0) == region)) { 1765 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1766 } 1767 if (address->is_Phi() && address->in(0) == region) { 1768 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1769 } 1770 if (base_is_phi && (base->in(0) == region)) { 1771 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1772 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1773 x->set_req(Address, adr_x); 1774 } 1775 } 1776 // Check for a 'win' on some paths 1777 const Type *t = x->Value(igvn); 1778 1779 bool singleton = t->singleton(); 1780 1781 // See comments in PhaseIdealLoop::split_thru_phi(). 1782 if (singleton && t == Type::TOP) { 1783 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1784 } 1785 1786 if (singleton) { 1787 x = igvn->makecon(t); 1788 } else { 1789 // We now call Identity to try to simplify the cloned node. 1790 // Note that some Identity methods call phase->type(this). 1791 // Make sure that the type array is big enough for 1792 // our new node, even though we may throw the node away. 1793 // (This tweaking with igvn only works because x is a new node.) 1794 igvn->set_type(x, t); 1795 // If x is a TypeNode, capture any more-precise type permanently into Node 1796 // otherwise it will be not updated during igvn->transform since 1797 // igvn->type(x) is set to x->Value() already. 1798 x->raise_bottom_type(t); 1799 Node* y = x->Identity(igvn); 1800 if (y != x) { 1801 x = y; 1802 } else { 1803 y = igvn->hash_find_insert(x); 1804 if (y) { 1805 x = y; 1806 } else { 1807 // Else x is a new node we are keeping 1808 // We do not need register_new_node_with_optimizer 1809 // because set_type has already been called. 1810 igvn->_worklist.push(x); 1811 } 1812 } 1813 } 1814 if (x != the_clone && the_clone != nullptr) { 1815 igvn->remove_dead_node(the_clone); 1816 } 1817 phi->set_req(i, x); 1818 } 1819 // Record Phi 1820 igvn->register_new_node_with_optimizer(phi); 1821 return phi; 1822 } 1823 1824 AllocateNode* LoadNode::is_new_object_mark_load() const { 1825 if (Opcode() == Op_LoadX) { 1826 Node* address = in(MemNode::Address); 1827 AllocateNode* alloc = AllocateNode::Ideal_allocation(address); 1828 Node* mem = in(MemNode::Memory); 1829 if (alloc != nullptr && mem->is_Proj() && 1830 mem->in(0) != nullptr && 1831 mem->in(0) == alloc->initialization() && 1832 alloc->initialization()->proj_out_or_null(0) != nullptr) { 1833 return alloc; 1834 } 1835 } 1836 return nullptr; 1837 } 1838 1839 1840 //------------------------------Ideal------------------------------------------ 1841 // If the load is from Field memory and the pointer is non-null, it might be possible to 1842 // zero out the control input. 1843 // If the offset is constant and the base is an object allocation, 1844 // try to hook me up to the exact initializing store. 1845 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1846 if (has_pinned_control_dependency()) { 1847 return nullptr; 1848 } 1849 Node* p = MemNode::Ideal_common(phase, can_reshape); 1850 if (p) return (p == NodeSentinel) ? nullptr : p; 1851 1852 Node* ctrl = in(MemNode::Control); 1853 Node* address = in(MemNode::Address); 1854 bool progress = false; 1855 1856 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) && 1857 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes()); 1858 1859 // Skip up past a SafePoint control. Cannot do this for Stores because 1860 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1861 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint && 1862 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw && 1863 !addr_mark && 1864 (depends_only_on_test() || has_unknown_control_dependency())) { 1865 ctrl = ctrl->in(0); 1866 set_req(MemNode::Control,ctrl); 1867 progress = true; 1868 } 1869 1870 intptr_t ignore = 0; 1871 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1872 if (base != nullptr 1873 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1874 // Check for useless control edge in some common special cases 1875 if (in(MemNode::Control) != nullptr 1876 && can_remove_control() 1877 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1878 && all_controls_dominate(base, phase->C->start())) { 1879 // A method-invariant, non-null address (constant or 'this' argument). 1880 set_req(MemNode::Control, nullptr); 1881 progress = true; 1882 } 1883 } 1884 1885 Node* mem = in(MemNode::Memory); 1886 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1887 1888 if (can_reshape && (addr_t != nullptr)) { 1889 // try to optimize our memory input 1890 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1891 if (opt_mem != mem) { 1892 set_req_X(MemNode::Memory, opt_mem, phase); 1893 if (phase->type( opt_mem ) == Type::TOP) return nullptr; 1894 return this; 1895 } 1896 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1897 if ((t_oop != nullptr) && 1898 (t_oop->is_known_instance_field() || 1899 t_oop->is_ptr_to_boxed_value())) { 1900 PhaseIterGVN *igvn = phase->is_IterGVN(); 1901 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true"); 1902 if (igvn->_worklist.member(opt_mem)) { 1903 // Delay this transformation until memory Phi is processed. 1904 igvn->_worklist.push(this); 1905 return nullptr; 1906 } 1907 // Split instance field load through Phi. 1908 Node* result = split_through_phi(phase); 1909 if (result != nullptr) return result; 1910 1911 if (t_oop->is_ptr_to_boxed_value()) { 1912 Node* result = eliminate_autobox(igvn); 1913 if (result != nullptr) return result; 1914 } 1915 } 1916 } 1917 1918 // Is there a dominating load that loads the same value? Leave 1919 // anything that is not a load of a field/array element (like 1920 // barriers etc.) alone 1921 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) { 1922 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1923 Node *use = mem->fast_out(i); 1924 if (use != this && 1925 use->Opcode() == Opcode() && 1926 use->in(0) != nullptr && 1927 use->in(0) != in(0) && 1928 use->in(Address) == in(Address)) { 1929 Node* ctl = in(0); 1930 for (int i = 0; i < 10 && ctl != nullptr; i++) { 1931 ctl = IfNode::up_one_dom(ctl); 1932 if (ctl == use->in(0)) { 1933 set_req(0, use->in(0)); 1934 return this; 1935 } 1936 } 1937 } 1938 } 1939 } 1940 1941 // Check for prior store with a different base or offset; make Load 1942 // independent. Skip through any number of them. Bail out if the stores 1943 // are in an endless dead cycle and report no progress. This is a key 1944 // transform for Reflection. However, if after skipping through the Stores 1945 // we can't then fold up against a prior store do NOT do the transform as 1946 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1947 // array memory alive twice: once for the hoisted Load and again after the 1948 // bypassed Store. This situation only works if EVERYBODY who does 1949 // anti-dependence work knows how to bypass. I.e. we need all 1950 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1951 // the alias index stuff. So instead, peek through Stores and IFF we can 1952 // fold up, do so. 1953 Node* prev_mem = find_previous_store(phase); 1954 if (prev_mem != nullptr) { 1955 Node* value = can_see_arraycopy_value(prev_mem, phase); 1956 if (value != nullptr) { 1957 return value; 1958 } 1959 } 1960 // Steps (a), (b): Walk past independent stores to find an exact match. 1961 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) { 1962 // (c) See if we can fold up on the spot, but don't fold up here. 1963 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1964 // just return a prior value, which is done by Identity calls. 1965 if (can_see_stored_value(prev_mem, phase)) { 1966 // Make ready for step (d): 1967 set_req_X(MemNode::Memory, prev_mem, phase); 1968 return this; 1969 } 1970 } 1971 1972 return progress ? this : nullptr; 1973 } 1974 1975 // Helper to recognize certain Klass fields which are invariant across 1976 // some group of array types (e.g., int[] or all T[] where T < Object). 1977 const Type* 1978 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1979 ciKlass* klass) const { 1980 assert(!UseCompactObjectHeaders || tkls->offset() != in_bytes(Klass::prototype_header_offset()), 1981 "must not happen"); 1982 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1983 // The field is Klass::_access_flags. Return its (constant) value. 1984 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1985 assert(Opcode() == Op_LoadUS, "must load an unsigned short from _access_flags"); 1986 return TypeInt::make(klass->access_flags()); 1987 } 1988 if (tkls->offset() == in_bytes(Klass::misc_flags_offset())) { 1989 // The field is Klass::_misc_flags. Return its (constant) value. 1990 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1991 assert(Opcode() == Op_LoadUB, "must load an unsigned byte from _misc_flags"); 1992 return TypeInt::make(klass->misc_flags()); 1993 } 1994 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1995 // The field is Klass::_layout_helper. Return its constant value if known. 1996 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1997 return TypeInt::make(klass->layout_helper()); 1998 } 1999 2000 // No match. 2001 return nullptr; 2002 } 2003 2004 //------------------------------Value----------------------------------------- 2005 const Type* LoadNode::Value(PhaseGVN* phase) const { 2006 // Either input is TOP ==> the result is TOP 2007 Node* mem = in(MemNode::Memory); 2008 const Type *t1 = phase->type(mem); 2009 if (t1 == Type::TOP) return Type::TOP; 2010 Node* adr = in(MemNode::Address); 2011 const TypePtr* tp = phase->type(adr)->isa_ptr(); 2012 if (tp == nullptr || tp->empty()) return Type::TOP; 2013 int off = tp->offset(); 2014 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 2015 Compile* C = phase->C; 2016 2017 // If we are loading from a freshly-allocated object, produce a zero, 2018 // if the load is provably beyond the header of the object. 2019 // (Also allow a variable load from a fresh array to produce zero.) 2020 const TypeOopPtr* tinst = tp->isa_oopptr(); 2021 bool is_instance = (tinst != nullptr) && tinst->is_known_instance_field(); 2022 Node* value = can_see_stored_value(mem, phase); 2023 if (value != nullptr && value->is_Con()) { 2024 assert(value->bottom_type()->higher_equal(_type), "sanity"); 2025 return value->bottom_type(); 2026 } 2027 2028 // Try to guess loaded type from pointer type 2029 if (tp->isa_aryptr()) { 2030 const TypeAryPtr* ary = tp->is_aryptr(); 2031 const Type* t = ary->elem(); 2032 2033 // Determine whether the reference is beyond the header or not, by comparing 2034 // the offset against the offset of the start of the array's data. 2035 // Different array types begin at slightly different offsets (12 vs. 16). 2036 // We choose T_BYTE as an example base type that is least restrictive 2037 // as to alignment, which will therefore produce the smallest 2038 // possible base offset. 2039 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 2040 const bool off_beyond_header = (off >= min_base_off); 2041 2042 // Try to constant-fold a stable array element. 2043 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) { 2044 // Make sure the reference is not into the header and the offset is constant 2045 ciObject* aobj = ary->const_oop(); 2046 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 2047 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0); 2048 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, 2049 stable_dimension, 2050 value_basic_type(), is_unsigned()); 2051 if (con_type != nullptr) { 2052 return con_type; 2053 } 2054 } 2055 } 2056 2057 // Don't do this for integer types. There is only potential profit if 2058 // the element type t is lower than _type; that is, for int types, if _type is 2059 // more restrictive than t. This only happens here if one is short and the other 2060 // char (both 16 bits), and in those cases we've made an intentional decision 2061 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 2062 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 2063 // 2064 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 2065 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 2066 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 2067 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 2068 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 2069 // In fact, that could have been the original type of p1, and p1 could have 2070 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 2071 // expression (LShiftL quux 3) independently optimized to the constant 8. 2072 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr) 2073 && (_type->isa_vect() == nullptr) 2074 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 2075 // t might actually be lower than _type, if _type is a unique 2076 // concrete subclass of abstract class t. 2077 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header? 2078 const Type* jt = t->join_speculative(_type); 2079 // In any case, do not allow the join, per se, to empty out the type. 2080 if (jt->empty() && !t->empty()) { 2081 // This can happen if a interface-typed array narrows to a class type. 2082 jt = _type; 2083 } 2084 #ifdef ASSERT 2085 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 2086 // The pointers in the autobox arrays are always non-null 2087 Node* base = adr->in(AddPNode::Base); 2088 if ((base != nullptr) && base->is_DecodeN()) { 2089 // Get LoadN node which loads IntegerCache.cache field 2090 base = base->in(1); 2091 } 2092 if ((base != nullptr) && base->is_Con()) { 2093 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 2094 if ((base_type != nullptr) && base_type->is_autobox_cache()) { 2095 // It could be narrow oop 2096 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 2097 } 2098 } 2099 } 2100 #endif 2101 return jt; 2102 } 2103 } 2104 } else if (tp->base() == Type::InstPtr) { 2105 assert( off != Type::OffsetBot || 2106 // arrays can be cast to Objects 2107 !tp->isa_instptr() || 2108 tp->is_instptr()->instance_klass()->is_java_lang_Object() || 2109 // unsafe field access may not have a constant offset 2110 C->has_unsafe_access(), 2111 "Field accesses must be precise" ); 2112 // For oop loads, we expect the _type to be precise. 2113 2114 // Optimize loads from constant fields. 2115 const TypeInstPtr* tinst = tp->is_instptr(); 2116 ciObject* const_oop = tinst->const_oop(); 2117 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) { 2118 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), value_basic_type()); 2119 if (con_type != nullptr) { 2120 return con_type; 2121 } 2122 } 2123 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) { 2124 assert(off != Type::OffsetBot || 2125 !tp->isa_instklassptr() || 2126 // arrays can be cast to Objects 2127 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() || 2128 // also allow array-loading from the primary supertype 2129 // array during subtype checks 2130 Opcode() == Op_LoadKlass, 2131 "Field accesses must be precise"); 2132 // For klass/static loads, we expect the _type to be precise 2133 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) { 2134 /* With mirrors being an indirect in the Klass* 2135 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset)) 2136 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass). 2137 * 2138 * So check the type and klass of the node before the LoadP. 2139 */ 2140 Node* adr2 = adr->in(MemNode::Address); 2141 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2142 if (tkls != nullptr && !StressReflectiveCode) { 2143 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 2144 ciKlass* klass = tkls->exact_klass(); 2145 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 2146 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 2147 return TypeInstPtr::make(klass->java_mirror()); 2148 } 2149 } 2150 } 2151 2152 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2153 if (tkls != nullptr) { 2154 if (tkls->is_loaded() && tkls->klass_is_exact()) { 2155 ciKlass* klass = tkls->exact_klass(); 2156 // We are loading a field from a Klass metaobject whose identity 2157 // is known at compile time (the type is "exact" or "precise"). 2158 // Check for fields we know are maintained as constants by the VM. 2159 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 2160 // The field is Klass::_super_check_offset. Return its (constant) value. 2161 // (Folds up type checking code.) 2162 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 2163 return TypeInt::make(klass->super_check_offset()); 2164 } 2165 if (UseCompactObjectHeaders) { 2166 if (tkls->offset() == in_bytes(Klass::prototype_header_offset())) { 2167 // The field is Klass::_prototype_header. Return its (constant) value. 2168 assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header"); 2169 return TypeX::make(klass->prototype_header()); 2170 } 2171 } 2172 // Compute index into primary_supers array 2173 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 2174 // Check for overflowing; use unsigned compare to handle the negative case. 2175 if( depth < ciKlass::primary_super_limit() ) { 2176 // The field is an element of Klass::_primary_supers. Return its (constant) value. 2177 // (Folds up type checking code.) 2178 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 2179 ciKlass *ss = klass->super_of_depth(depth); 2180 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR; 2181 } 2182 const Type* aift = load_array_final_field(tkls, klass); 2183 if (aift != nullptr) return aift; 2184 } 2185 2186 // We can still check if we are loading from the primary_supers array at a 2187 // shallow enough depth. Even though the klass is not exact, entries less 2188 // than or equal to its super depth are correct. 2189 if (tkls->is_loaded()) { 2190 ciKlass* klass = nullptr; 2191 if (tkls->isa_instklassptr()) { 2192 klass = tkls->is_instklassptr()->instance_klass(); 2193 } else { 2194 int dims; 2195 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims); 2196 if (inner->isa_instklassptr()) { 2197 klass = inner->is_instklassptr()->instance_klass(); 2198 klass = ciObjArrayKlass::make(klass, dims); 2199 } 2200 } 2201 if (klass != nullptr) { 2202 // Compute index into primary_supers array 2203 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 2204 // Check for overflowing; use unsigned compare to handle the negative case. 2205 if (depth < ciKlass::primary_super_limit() && 2206 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case 2207 // The field is an element of Klass::_primary_supers. Return its (constant) value. 2208 // (Folds up type checking code.) 2209 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 2210 ciKlass *ss = klass->super_of_depth(depth); 2211 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR; 2212 } 2213 } 2214 } 2215 2216 // If the type is enough to determine that the thing is not an array, 2217 // we can give the layout_helper a positive interval type. 2218 // This will help short-circuit some reflective code. 2219 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) && 2220 tkls->isa_instklassptr() && // not directly typed as an array 2221 !tkls->is_instklassptr()->might_be_an_array() // not the supertype of all T[] (java.lang.Object) or has an interface that is not Serializable or Cloneable 2222 ) { 2223 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 2224 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 2225 // The key property of this type is that it folds up tests 2226 // for array-ness, since it proves that the layout_helper is positive. 2227 // Thus, a generic value like the basic object layout helper works fine. 2228 return TypeInt::make(min_size, max_jint, Type::WidenMin); 2229 } 2230 } 2231 2232 bool is_vect = (_type->isa_vect() != nullptr); 2233 if (is_instance && !is_vect) { 2234 // If we have an instance type and our memory input is the 2235 // programs's initial memory state, there is no matching store, 2236 // so just return a zero of the appropriate type - 2237 // except if it is vectorized - then we have no zero constant. 2238 Node *mem = in(MemNode::Memory); 2239 if (mem->is_Parm() && mem->in(0)->is_Start()) { 2240 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 2241 return Type::get_zero_type(_type->basic_type()); 2242 } 2243 } 2244 2245 if (!UseCompactObjectHeaders) { 2246 Node* alloc = is_new_object_mark_load(); 2247 if (alloc != nullptr) { 2248 return TypeX::make(markWord::prototype().value()); 2249 } 2250 } 2251 2252 return _type; 2253 } 2254 2255 //------------------------------match_edge------------------------------------- 2256 // Do we Match on this edge index or not? Match only the address. 2257 uint LoadNode::match_edge(uint idx) const { 2258 return idx == MemNode::Address; 2259 } 2260 2261 //--------------------------LoadBNode::Ideal-------------------------------------- 2262 // 2263 // If the previous store is to the same address as this load, 2264 // and the value stored was larger than a byte, replace this load 2265 // with the value stored truncated to a byte. If no truncation is 2266 // needed, the replacement is done in LoadNode::Identity(). 2267 // 2268 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2269 Node* mem = in(MemNode::Memory); 2270 Node* value = can_see_stored_value(mem,phase); 2271 if (value != nullptr) { 2272 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false); 2273 if (narrow != value) { 2274 return narrow; 2275 } 2276 } 2277 // Identity call will handle the case where truncation is not needed. 2278 return LoadNode::Ideal(phase, can_reshape); 2279 } 2280 2281 const Type* LoadBNode::Value(PhaseGVN* phase) const { 2282 Node* mem = in(MemNode::Memory); 2283 Node* value = can_see_stored_value(mem,phase); 2284 if (value != nullptr && value->is_Con() && 2285 !value->bottom_type()->higher_equal(_type)) { 2286 // If the input to the store does not fit with the load's result type, 2287 // it must be truncated. We can't delay until Ideal call since 2288 // a singleton Value is needed for split_thru_phi optimization. 2289 int con = value->get_int(); 2290 return TypeInt::make((con << 24) >> 24); 2291 } 2292 return LoadNode::Value(phase); 2293 } 2294 2295 //--------------------------LoadUBNode::Ideal------------------------------------- 2296 // 2297 // If the previous store is to the same address as this load, 2298 // and the value stored was larger than a byte, replace this load 2299 // with the value stored truncated to a byte. If no truncation is 2300 // needed, the replacement is done in LoadNode::Identity(). 2301 // 2302 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2303 Node* mem = in(MemNode::Memory); 2304 Node* value = can_see_stored_value(mem, phase); 2305 if (value != nullptr) { 2306 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false); 2307 if (narrow != value) { 2308 return narrow; 2309 } 2310 } 2311 // Identity call will handle the case where truncation is not needed. 2312 return LoadNode::Ideal(phase, can_reshape); 2313 } 2314 2315 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 2316 Node* mem = in(MemNode::Memory); 2317 Node* value = can_see_stored_value(mem,phase); 2318 if (value != nullptr && value->is_Con() && 2319 !value->bottom_type()->higher_equal(_type)) { 2320 // If the input to the store does not fit with the load's result type, 2321 // it must be truncated. We can't delay until Ideal call since 2322 // a singleton Value is needed for split_thru_phi optimization. 2323 int con = value->get_int(); 2324 return TypeInt::make(con & 0xFF); 2325 } 2326 return LoadNode::Value(phase); 2327 } 2328 2329 //--------------------------LoadUSNode::Ideal------------------------------------- 2330 // 2331 // If the previous store is to the same address as this load, 2332 // and the value stored was larger than a char, replace this load 2333 // with the value stored truncated to a char. If no truncation is 2334 // needed, the replacement is done in LoadNode::Identity(). 2335 // 2336 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2337 Node* mem = in(MemNode::Memory); 2338 Node* value = can_see_stored_value(mem,phase); 2339 if (value != nullptr) { 2340 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false); 2341 if (narrow != value) { 2342 return narrow; 2343 } 2344 } 2345 // Identity call will handle the case where truncation is not needed. 2346 return LoadNode::Ideal(phase, can_reshape); 2347 } 2348 2349 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 2350 Node* mem = in(MemNode::Memory); 2351 Node* value = can_see_stored_value(mem,phase); 2352 if (value != nullptr && value->is_Con() && 2353 !value->bottom_type()->higher_equal(_type)) { 2354 // If the input to the store does not fit with the load's result type, 2355 // it must be truncated. We can't delay until Ideal call since 2356 // a singleton Value is needed for split_thru_phi optimization. 2357 int con = value->get_int(); 2358 return TypeInt::make(con & 0xFFFF); 2359 } 2360 return LoadNode::Value(phase); 2361 } 2362 2363 //--------------------------LoadSNode::Ideal-------------------------------------- 2364 // 2365 // If the previous store is to the same address as this load, 2366 // and the value stored was larger than a short, replace this load 2367 // with the value stored truncated to a short. If no truncation is 2368 // needed, the replacement is done in LoadNode::Identity(). 2369 // 2370 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2371 Node* mem = in(MemNode::Memory); 2372 Node* value = can_see_stored_value(mem,phase); 2373 if (value != nullptr) { 2374 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false); 2375 if (narrow != value) { 2376 return narrow; 2377 } 2378 } 2379 // Identity call will handle the case where truncation is not needed. 2380 return LoadNode::Ideal(phase, can_reshape); 2381 } 2382 2383 const Type* LoadSNode::Value(PhaseGVN* phase) const { 2384 Node* mem = in(MemNode::Memory); 2385 Node* value = can_see_stored_value(mem,phase); 2386 if (value != nullptr && value->is_Con() && 2387 !value->bottom_type()->higher_equal(_type)) { 2388 // If the input to the store does not fit with the load's result type, 2389 // it must be truncated. We can't delay until Ideal call since 2390 // a singleton Value is needed for split_thru_phi optimization. 2391 int con = value->get_int(); 2392 return TypeInt::make((con << 16) >> 16); 2393 } 2394 return LoadNode::Value(phase); 2395 } 2396 2397 //============================================================================= 2398 //----------------------------LoadKlassNode::make------------------------------ 2399 // Polymorphic factory method: 2400 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 2401 // sanity check the alias category against the created node type 2402 const TypePtr* adr_type = adr->bottom_type()->isa_ptr(); 2403 assert(adr_type != nullptr, "expecting TypeKlassPtr"); 2404 #ifdef _LP64 2405 if (adr_type->is_ptr_to_narrowklass()) { 2406 assert(UseCompressedClassPointers, "no compressed klasses"); 2407 Node* load_klass = gvn.transform(new LoadNKlassNode(mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2408 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2409 } 2410 #endif 2411 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2412 return new LoadKlassNode(mem, adr, at, tk, MemNode::unordered); 2413 } 2414 2415 //------------------------------Value------------------------------------------ 2416 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 2417 return klass_value_common(phase); 2418 } 2419 2420 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2421 // Either input is TOP ==> the result is TOP 2422 const Type *t1 = phase->type( in(MemNode::Memory) ); 2423 if (t1 == Type::TOP) return Type::TOP; 2424 Node *adr = in(MemNode::Address); 2425 const Type *t2 = phase->type( adr ); 2426 if (t2 == Type::TOP) return Type::TOP; 2427 const TypePtr *tp = t2->is_ptr(); 2428 if (TypePtr::above_centerline(tp->ptr()) || 2429 tp->ptr() == TypePtr::Null) return Type::TOP; 2430 2431 // Return a more precise klass, if possible 2432 const TypeInstPtr *tinst = tp->isa_instptr(); 2433 if (tinst != nullptr) { 2434 ciInstanceKlass* ik = tinst->instance_klass(); 2435 int offset = tinst->offset(); 2436 if (ik == phase->C->env()->Class_klass() 2437 && (offset == java_lang_Class::klass_offset() || 2438 offset == java_lang_Class::array_klass_offset())) { 2439 // We are loading a special hidden field from a Class mirror object, 2440 // the field which points to the VM's Klass metaobject. 2441 ciType* t = tinst->java_mirror_type(); 2442 // java_mirror_type returns non-null for compile-time Class constants. 2443 if (t != nullptr) { 2444 // constant oop => constant klass 2445 if (offset == java_lang_Class::array_klass_offset()) { 2446 if (t->is_void()) { 2447 // We cannot create a void array. Since void is a primitive type return null 2448 // klass. Users of this result need to do a null check on the returned klass. 2449 return TypePtr::NULL_PTR; 2450 } 2451 return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces); 2452 } 2453 if (!t->is_klass()) { 2454 // a primitive Class (e.g., int.class) has null for a klass field 2455 return TypePtr::NULL_PTR; 2456 } 2457 // Fold up the load of the hidden field 2458 return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces); 2459 } 2460 // non-constant mirror, so we can't tell what's going on 2461 } 2462 if (!tinst->is_loaded()) 2463 return _type; // Bail out if not loaded 2464 if (offset == oopDesc::klass_offset_in_bytes()) { 2465 return tinst->as_klass_type(true); 2466 } 2467 } 2468 2469 // Check for loading klass from an array 2470 const TypeAryPtr *tary = tp->isa_aryptr(); 2471 if (tary != nullptr && 2472 tary->offset() == oopDesc::klass_offset_in_bytes()) { 2473 return tary->as_klass_type(true); 2474 } 2475 2476 // Check for loading klass from an array klass 2477 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2478 if (tkls != nullptr && !StressReflectiveCode) { 2479 if (!tkls->is_loaded()) 2480 return _type; // Bail out if not loaded 2481 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() && 2482 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2483 // // Always returning precise element type is incorrect, 2484 // // e.g., element type could be object and array may contain strings 2485 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2486 2487 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2488 // according to the element type's subclassing. 2489 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact()); 2490 } 2491 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() && 2492 tkls->offset() == in_bytes(Klass::super_offset())) { 2493 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super(); 2494 // The field is Klass::_super. Return its (constant) value. 2495 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2496 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR; 2497 } 2498 } 2499 2500 if (tkls != nullptr && !UseSecondarySupersCache 2501 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) { 2502 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled. 2503 return TypePtr::NULL_PTR; 2504 } 2505 2506 // Bailout case 2507 return LoadNode::Value(phase); 2508 } 2509 2510 //------------------------------Identity--------------------------------------- 2511 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2512 // Also feed through the klass in Allocate(...klass...)._klass. 2513 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2514 return klass_identity_common(phase); 2515 } 2516 2517 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2518 Node* x = LoadNode::Identity(phase); 2519 if (x != this) return x; 2520 2521 // Take apart the address into an oop and offset. 2522 // Return 'this' if we cannot. 2523 Node* adr = in(MemNode::Address); 2524 intptr_t offset = 0; 2525 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2526 if (base == nullptr) return this; 2527 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2528 if (toop == nullptr) return this; 2529 2530 // Step over potential GC barrier for OopHandle resolve 2531 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2532 if (bs->is_gc_barrier_node(base)) { 2533 base = bs->step_over_gc_barrier(base); 2534 } 2535 2536 // We can fetch the klass directly through an AllocateNode. 2537 // This works even if the klass is not constant (clone or newArray). 2538 if (offset == oopDesc::klass_offset_in_bytes()) { 2539 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2540 if (allocated_klass != nullptr) { 2541 return allocated_klass; 2542 } 2543 } 2544 2545 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2546 // See inline_native_Class_query for occurrences of these patterns. 2547 // Java Example: x.getClass().isAssignableFrom(y) 2548 // 2549 // This improves reflective code, often making the Class 2550 // mirror go completely dead. (Current exception: Class 2551 // mirrors may appear in debug info, but we could clean them out by 2552 // introducing a new debug info operator for Klass.java_mirror). 2553 2554 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass() 2555 && offset == java_lang_Class::klass_offset()) { 2556 if (base->is_Load()) { 2557 Node* base2 = base->in(MemNode::Address); 2558 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */ 2559 Node* adr2 = base2->in(MemNode::Address); 2560 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2561 if (tkls != nullptr && !tkls->empty() 2562 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr()) 2563 && adr2->is_AddP() 2564 ) { 2565 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2566 if (tkls->offset() == mirror_field) { 2567 return adr2->in(AddPNode::Base); 2568 } 2569 } 2570 } 2571 } 2572 } 2573 2574 return this; 2575 } 2576 2577 LoadNode* LoadNode::clone_pinned() const { 2578 LoadNode* ld = clone()->as_Load(); 2579 ld->_control_dependency = UnknownControl; 2580 return ld; 2581 } 2582 2583 2584 //------------------------------Value------------------------------------------ 2585 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2586 const Type *t = klass_value_common(phase); 2587 if (t == Type::TOP) 2588 return t; 2589 2590 return t->make_narrowklass(); 2591 } 2592 2593 //------------------------------Identity--------------------------------------- 2594 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2595 // Also feed through the klass in Allocate(...klass...)._klass. 2596 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2597 Node *x = klass_identity_common(phase); 2598 2599 const Type *t = phase->type( x ); 2600 if( t == Type::TOP ) return x; 2601 if( t->isa_narrowklass()) return x; 2602 assert (!t->isa_narrowoop(), "no narrow oop here"); 2603 2604 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2605 } 2606 2607 //------------------------------Value----------------------------------------- 2608 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2609 // Either input is TOP ==> the result is TOP 2610 const Type *t1 = phase->type( in(MemNode::Memory) ); 2611 if( t1 == Type::TOP ) return Type::TOP; 2612 Node *adr = in(MemNode::Address); 2613 const Type *t2 = phase->type( adr ); 2614 if( t2 == Type::TOP ) return Type::TOP; 2615 const TypePtr *tp = t2->is_ptr(); 2616 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2617 const TypeAryPtr *tap = tp->isa_aryptr(); 2618 if( !tap ) return _type; 2619 return tap->size(); 2620 } 2621 2622 //-------------------------------Ideal--------------------------------------- 2623 // Feed through the length in AllocateArray(...length...)._length. 2624 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2625 Node* p = MemNode::Ideal_common(phase, can_reshape); 2626 if (p) return (p == NodeSentinel) ? nullptr : p; 2627 2628 // Take apart the address into an oop and offset. 2629 // Return 'this' if we cannot. 2630 Node* adr = in(MemNode::Address); 2631 intptr_t offset = 0; 2632 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2633 if (base == nullptr) return nullptr; 2634 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2635 if (tary == nullptr) return nullptr; 2636 2637 // We can fetch the length directly through an AllocateArrayNode. 2638 // This works even if the length is not constant (clone or newArray). 2639 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2640 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base); 2641 if (alloc != nullptr) { 2642 Node* allocated_length = alloc->Ideal_length(); 2643 Node* len = alloc->make_ideal_length(tary, phase); 2644 if (allocated_length != len) { 2645 // New CastII improves on this. 2646 return len; 2647 } 2648 } 2649 } 2650 2651 return nullptr; 2652 } 2653 2654 //------------------------------Identity--------------------------------------- 2655 // Feed through the length in AllocateArray(...length...)._length. 2656 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2657 Node* x = LoadINode::Identity(phase); 2658 if (x != this) return x; 2659 2660 // Take apart the address into an oop and offset. 2661 // Return 'this' if we cannot. 2662 Node* adr = in(MemNode::Address); 2663 intptr_t offset = 0; 2664 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2665 if (base == nullptr) return this; 2666 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2667 if (tary == nullptr) return this; 2668 2669 // We can fetch the length directly through an AllocateArrayNode. 2670 // This works even if the length is not constant (clone or newArray). 2671 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2672 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base); 2673 if (alloc != nullptr) { 2674 Node* allocated_length = alloc->Ideal_length(); 2675 // Do not allow make_ideal_length to allocate a CastII node. 2676 Node* len = alloc->make_ideal_length(tary, phase, false); 2677 if (allocated_length == len) { 2678 // Return allocated_length only if it would not be improved by a CastII. 2679 return allocated_length; 2680 } 2681 } 2682 } 2683 2684 return this; 2685 2686 } 2687 2688 //============================================================================= 2689 //---------------------------StoreNode::make----------------------------------- 2690 // Polymorphic factory method: 2691 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo, bool require_atomic_access) { 2692 assert((mo == unordered || mo == release), "unexpected"); 2693 Compile* C = gvn.C; 2694 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2695 ctl != nullptr, "raw memory operations should have control edge"); 2696 2697 switch (bt) { 2698 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2699 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2700 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2701 case T_CHAR: 2702 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2703 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access); 2704 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2705 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access); 2706 case T_METADATA: 2707 case T_ADDRESS: 2708 case T_OBJECT: 2709 #ifdef _LP64 2710 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2711 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2712 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2713 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2714 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2715 adr->bottom_type()->isa_rawptr())) { 2716 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2717 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2718 } 2719 #endif 2720 { 2721 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2722 } 2723 default: 2724 ShouldNotReachHere(); 2725 return (StoreNode*)nullptr; 2726 } 2727 } 2728 2729 //--------------------------bottom_type---------------------------------------- 2730 const Type *StoreNode::bottom_type() const { 2731 return Type::MEMORY; 2732 } 2733 2734 //------------------------------hash------------------------------------------- 2735 uint StoreNode::hash() const { 2736 // unroll addition of interesting fields 2737 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2738 2739 // Since they are not commoned, do not hash them: 2740 return NO_HASH; 2741 } 2742 2743 // Link together multiple stores (B/S/C/I) into a longer one. 2744 // 2745 // Example: _store = StoreB[i+3] 2746 // 2747 // RangeCheck[i+0] RangeCheck[i+0] 2748 // StoreB[i+0] 2749 // RangeCheck[i+3] RangeCheck[i+3] 2750 // StoreB[i+1] --> pass: fail: 2751 // StoreB[i+2] StoreI[i+0] StoreB[i+0] 2752 // StoreB[i+3] 2753 // 2754 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before 2755 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After 2756 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the 2757 // failing path. 2758 // 2759 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with: 2760 // - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop, 2761 // and possibly no RangeChecks remain between the stores. 2762 // - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks, 2763 // and those later RangeChecks can be removed. Example: 2764 // 2765 // RangeCheck[i+0] RangeCheck[i+0] <- before first store 2766 // StoreB[i+0] StoreB[i+0] <- first store 2767 // RangeCheck[i+1] --> smeared --> RangeCheck[i+3] <- only RC between first and last store 2768 // StoreB[i+1] StoreB[i+1] <- second store 2769 // RangeCheck[i+2] --> removed 2770 // StoreB[i+2] StoreB[i+2] 2771 // RangeCheck[i+3] --> removed 2772 // StoreB[i+3] StoreB[i+3] <- last store 2773 // 2774 // Thus, it is a common pattern that between the first and last store in a chain 2775 // of adjacent stores there remains exactly one RangeCheck, located between the 2776 // first and the second store (e.g. RangeCheck[i+3]). 2777 // 2778 class MergePrimitiveStores : public StackObj { 2779 private: 2780 PhaseGVN* const _phase; 2781 StoreNode* const _store; 2782 // State machine with initial state Unknown 2783 // Allowed transitions: 2784 // Unknown -> Const 2785 // Unknown -> Platform 2786 // Unknown -> Reverse 2787 // Unknown -> NotAdjacent 2788 // Const -> Const 2789 // Const -> NotAdjacent 2790 // Platform -> Platform 2791 // Platform -> NotAdjacent 2792 // Reverse -> Reverse 2793 // Reverse -> NotAdjacent 2794 // NotAdjacent -> NotAdjacent 2795 enum ValueOrder : uint8_t { 2796 Unknown, // Initial state 2797 Const, // Input values are const 2798 Platform, // Platform order 2799 Reverse, // Reverse platform order 2800 NotAdjacent // Not adjacent 2801 }; 2802 ValueOrder _value_order; 2803 2804 NOT_PRODUCT( const CHeapBitMap &_trace_tags; ) 2805 2806 public: 2807 MergePrimitiveStores(PhaseGVN* phase, StoreNode* store) : 2808 _phase(phase), _store(store), _value_order(ValueOrder::Unknown) 2809 NOT_PRODUCT( COMMA _trace_tags(Compile::current()->directive()->trace_merge_stores_tags()) ) 2810 {} 2811 2812 StoreNode* run(); 2813 2814 private: 2815 bool is_compatible_store(const StoreNode* other_store) const; 2816 bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const; 2817 bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const; 2818 static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase); 2819 enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure }; 2820 static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store); 2821 2822 class Status { 2823 private: 2824 StoreNode* _found_store; 2825 bool _found_range_check; 2826 2827 Status(StoreNode* found_store, bool found_range_check) 2828 : _found_store(found_store), _found_range_check(found_range_check) {} 2829 2830 public: 2831 StoreNode* found_store() const { return _found_store; } 2832 bool found_range_check() const { return _found_range_check; } 2833 static Status make_failure() { return Status(nullptr, false); } 2834 2835 static Status make(StoreNode* found_store, const CFGStatus cfg_status) { 2836 if (cfg_status == CFGStatus::Failure) { 2837 return Status::make_failure(); 2838 } 2839 return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck); 2840 } 2841 2842 #ifndef PRODUCT 2843 void print_on(outputStream* st) const { 2844 if (_found_store == nullptr) { 2845 st->print_cr("None"); 2846 } else { 2847 st->print_cr("Found[%d %s, %s]", _found_store->_idx, _found_store->Name(), 2848 _found_range_check ? "RC" : "no-RC"); 2849 } 2850 } 2851 #endif 2852 }; 2853 2854 enum ValueOrder find_adjacent_input_value_order(const Node* n1, const Node* n2, const int memory_size) const; 2855 Status find_adjacent_use_store(const StoreNode* def_store) const; 2856 Status find_adjacent_def_store(const StoreNode* use_store) const; 2857 Status find_use_store(const StoreNode* def_store) const; 2858 Status find_def_store(const StoreNode* use_store) const; 2859 Status find_use_store_unidirectional(const StoreNode* def_store) const; 2860 Status find_def_store_unidirectional(const StoreNode* use_store) const; 2861 2862 void collect_merge_list(Node_List& merge_list) const; 2863 Node* make_merged_input_value(const Node_List& merge_list); 2864 StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value); 2865 2866 #ifndef PRODUCT 2867 // Access to TraceMergeStores tags 2868 bool is_trace(TraceMergeStores::Tag tag) const { 2869 return _trace_tags.at(tag); 2870 } 2871 2872 bool is_trace_basic() const { 2873 return is_trace(TraceMergeStores::Tag::BASIC); 2874 } 2875 2876 bool is_trace_pointer_parsing() const { 2877 return is_trace(TraceMergeStores::Tag::POINTER_PARSING); 2878 } 2879 2880 bool is_trace_pointer_aliasing() const { 2881 return is_trace(TraceMergeStores::Tag::POINTER_ALIASING); 2882 } 2883 2884 bool is_trace_pointer_adjacency() const { 2885 return is_trace(TraceMergeStores::Tag::POINTER_ADJACENCY); 2886 } 2887 2888 bool is_trace_success() const { 2889 return is_trace(TraceMergeStores::Tag::SUCCESS); 2890 } 2891 #endif 2892 2893 NOT_PRODUCT( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; ) 2894 }; 2895 2896 StoreNode* MergePrimitiveStores::run() { 2897 // Check for B/S/C/I 2898 int opc = _store->Opcode(); 2899 if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) { 2900 return nullptr; 2901 } 2902 2903 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] MergePrimitiveStores::run: "); _store->dump(); }) 2904 2905 // The _store must be the "last" store in a chain. If we find a use we could merge with 2906 // then that use or a store further down is the "last" store. 2907 Status status_use = find_adjacent_use_store(_store); 2908 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect no use: "); status_use.print_on(tty); }) 2909 if (status_use.found_store() != nullptr) { 2910 return nullptr; 2911 } 2912 2913 // Check if we can merge with at least one def, so that we have at least 2 stores to merge. 2914 Status status_def = find_adjacent_def_store(_store); 2915 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect def: "); status_def.print_on(tty); }) 2916 Node* def_store = status_def.found_store(); 2917 if (def_store == nullptr) { 2918 return nullptr; 2919 } 2920 2921 // Initialize value order 2922 _value_order = find_adjacent_input_value_order(def_store->in(MemNode::ValueIn), 2923 _store->in(MemNode::ValueIn), 2924 _store->memory_size()); 2925 assert(_value_order != ValueOrder::NotAdjacent && _value_order != ValueOrder::Unknown, "Order should be checked"); 2926 2927 ResourceMark rm; 2928 Node_List merge_list; 2929 collect_merge_list(merge_list); 2930 2931 Node* merged_input_value = make_merged_input_value(merge_list); 2932 if (merged_input_value == nullptr) { return nullptr; } 2933 2934 StoreNode* merged_store = make_merged_store(merge_list, merged_input_value); 2935 2936 NOT_PRODUCT( if (is_trace_success()) { trace(merge_list, merged_input_value, merged_store); } ) 2937 2938 return merged_store; 2939 } 2940 2941 // Check compatibility between _store and other_store. 2942 bool MergePrimitiveStores::is_compatible_store(const StoreNode* other_store) const { 2943 int opc = _store->Opcode(); 2944 assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition"); 2945 2946 if (other_store == nullptr || 2947 _store->Opcode() != other_store->Opcode()) { 2948 return false; 2949 } 2950 2951 return true; 2952 } 2953 2954 bool MergePrimitiveStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const { 2955 if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn), 2956 use_store->in(MemNode::ValueIn), 2957 def_store->memory_size())) { 2958 return false; 2959 } 2960 2961 ResourceMark rm; 2962 #ifndef PRODUCT 2963 const TraceMemPointer trace(is_trace_pointer_parsing(), 2964 is_trace_pointer_aliasing(), 2965 is_trace_pointer_adjacency(), 2966 true); 2967 #endif 2968 const MemPointer pointer_use(use_store NOT_PRODUCT(COMMA trace)); 2969 const MemPointer pointer_def(def_store NOT_PRODUCT(COMMA trace)); 2970 return pointer_def.is_adjacent_to_and_before(pointer_use); 2971 } 2972 2973 // Check input values n1 and n2 can be merged and return the value order 2974 MergePrimitiveStores::ValueOrder MergePrimitiveStores::find_adjacent_input_value_order(const Node* n1, const Node* n2, 2975 const int memory_size) const { 2976 // Pattern: [n1 = ConI, n2 = ConI] 2977 if (n1->Opcode() == Op_ConI && n2->Opcode() == Op_ConI) { 2978 return ValueOrder::Const; 2979 } 2980 2981 Node const *base_n2; 2982 jint shift_n2; 2983 if (!is_con_RShift(n2, base_n2, shift_n2, _phase)) { 2984 return ValueOrder::NotAdjacent; 2985 } 2986 Node const *base_n1; 2987 jint shift_n1; 2988 if (!is_con_RShift(n1, base_n1, shift_n1, _phase)) { 2989 return ValueOrder::NotAdjacent; 2990 } 2991 2992 int bits_per_store = memory_size * 8; 2993 if (base_n1 != base_n2 || 2994 abs(shift_n1 - shift_n2) != bits_per_store || 2995 shift_n1 % bits_per_store != 0) { 2996 // Values are not adjacent 2997 return ValueOrder::NotAdjacent; 2998 } 2999 3000 // Detect value order 3001 #ifdef VM_LITTLE_ENDIAN 3002 return shift_n1 < shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)] 3003 : ValueOrder::Reverse; // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift] 3004 #else 3005 return shift_n1 > shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift] 3006 : ValueOrder::Reverse; // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)] 3007 #endif 3008 } 3009 3010 bool MergePrimitiveStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const { 3011 ValueOrder input_value_order = find_adjacent_input_value_order(n1, n2, memory_size); 3012 3013 switch (input_value_order) { 3014 case ValueOrder::NotAdjacent: 3015 return false; 3016 case ValueOrder::Reverse: 3017 if (memory_size != 1 || 3018 !Matcher::match_rule_supported(Op_ReverseBytesS) || 3019 !Matcher::match_rule_supported(Op_ReverseBytesI) || 3020 !Matcher::match_rule_supported(Op_ReverseBytesL)) { 3021 // ReverseBytes are not supported by platform 3022 return false; 3023 } 3024 // fall-through. 3025 case ValueOrder::Const: 3026 case ValueOrder::Platform: 3027 if (_value_order == ValueOrder::Unknown) { 3028 // Initial state is Unknown, and we find a valid input value order 3029 return true; 3030 } 3031 // The value order can not be changed 3032 return _value_order == input_value_order; 3033 case ValueOrder::Unknown: 3034 default: 3035 ShouldNotReachHere(); 3036 } 3037 return false; 3038 } 3039 3040 // Detect pattern: n = base_out >> shift_out 3041 bool MergePrimitiveStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase) { 3042 assert(n != nullptr, "precondition"); 3043 3044 int opc = n->Opcode(); 3045 if (opc == Op_ConvL2I) { 3046 n = n->in(1); 3047 opc = n->Opcode(); 3048 } 3049 3050 if ((opc == Op_RShiftI || 3051 opc == Op_RShiftL || 3052 opc == Op_URShiftI || 3053 opc == Op_URShiftL) && 3054 n->in(2)->is_ConI()) { 3055 base_out = n->in(1); 3056 shift_out = n->in(2)->get_int(); 3057 // The shift must be positive: 3058 return shift_out >= 0; 3059 } 3060 3061 if (phase->type(n)->isa_int() != nullptr || 3062 phase->type(n)->isa_long() != nullptr) { 3063 // (base >> 0) 3064 base_out = n; 3065 shift_out = 0; 3066 return true; 3067 } 3068 return false; 3069 } 3070 3071 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap. 3072 MergePrimitiveStores::CFGStatus MergePrimitiveStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) { 3073 assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship"); 3074 3075 Node* ctrl_use = use_store->in(MemNode::Control); 3076 Node* ctrl_def = def_store->in(MemNode::Control); 3077 if (ctrl_use == nullptr || ctrl_def == nullptr) { 3078 return CFGStatus::Failure; 3079 } 3080 3081 if (ctrl_use == ctrl_def) { 3082 // Same ctrl -> no RangeCheck in between. 3083 // Check: use_store must be the only use of def_store. 3084 if (def_store->outcnt() > 1) { 3085 return CFGStatus::Failure; 3086 } 3087 return CFGStatus::SuccessNoRangeCheck; 3088 } 3089 3090 // Different ctrl -> could have RangeCheck in between. 3091 // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and 3092 // 2. ctrl separated by RangeCheck. 3093 if (def_store->outcnt() != 2) { 3094 return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap. 3095 } 3096 int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1; 3097 Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem(); 3098 if (merge_mem == nullptr || 3099 merge_mem->outcnt() != 1) { 3100 return CFGStatus::Failure; // Does not have MergeMem for uncommon trap. 3101 } 3102 if (!ctrl_use->is_IfProj() || 3103 !ctrl_use->in(0)->is_RangeCheck() || 3104 ctrl_use->in(0)->outcnt() != 2) { 3105 return CFGStatus::Failure; // Not RangeCheck. 3106 } 3107 ProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj(); 3108 Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check); 3109 if (trap != merge_mem->unique_out() || 3110 ctrl_use->in(0)->in(0) != ctrl_def) { 3111 return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap. 3112 } 3113 3114 return CFGStatus::SuccessWithRangeCheck; 3115 } 3116 3117 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_use_store(const StoreNode* def_store) const { 3118 Status status_use = find_use_store(def_store); 3119 StoreNode* use_store = status_use.found_store(); 3120 if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) { 3121 return Status::make_failure(); 3122 } 3123 return status_use; 3124 } 3125 3126 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_def_store(const StoreNode* use_store) const { 3127 Status status_def = find_def_store(use_store); 3128 StoreNode* def_store = status_def.found_store(); 3129 if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) { 3130 return Status::make_failure(); 3131 } 3132 return status_def; 3133 } 3134 3135 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store(const StoreNode* def_store) const { 3136 Status status_use = find_use_store_unidirectional(def_store); 3137 3138 #ifdef ASSERT 3139 StoreNode* use_store = status_use.found_store(); 3140 if (use_store != nullptr) { 3141 Status status_def = find_def_store_unidirectional(use_store); 3142 assert(status_def.found_store() == def_store && 3143 status_def.found_range_check() == status_use.found_range_check(), 3144 "find_use_store and find_def_store must be symmetric"); 3145 } 3146 #endif 3147 3148 return status_use; 3149 } 3150 3151 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store(const StoreNode* use_store) const { 3152 Status status_def = find_def_store_unidirectional(use_store); 3153 3154 #ifdef ASSERT 3155 StoreNode* def_store = status_def.found_store(); 3156 if (def_store != nullptr) { 3157 Status status_use = find_use_store_unidirectional(def_store); 3158 assert(status_use.found_store() == use_store && 3159 status_use.found_range_check() == status_def.found_range_check(), 3160 "find_use_store and find_def_store must be symmetric"); 3161 } 3162 #endif 3163 3164 return status_def; 3165 } 3166 3167 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store_unidirectional(const StoreNode* def_store) const { 3168 assert(is_compatible_store(def_store), "precondition: must be compatible with _store"); 3169 3170 for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) { 3171 StoreNode* use_store = def_store->fast_out(i)->isa_Store(); 3172 if (is_compatible_store(use_store)) { 3173 return Status::make(use_store, cfg_status_for_pair(use_store, def_store)); 3174 } 3175 } 3176 3177 return Status::make_failure(); 3178 } 3179 3180 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store_unidirectional(const StoreNode* use_store) const { 3181 assert(is_compatible_store(use_store), "precondition: must be compatible with _store"); 3182 3183 StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store(); 3184 if (!is_compatible_store(def_store)) { 3185 return Status::make_failure(); 3186 } 3187 3188 return Status::make(def_store, cfg_status_for_pair(use_store, def_store)); 3189 } 3190 3191 void MergePrimitiveStores::collect_merge_list(Node_List& merge_list) const { 3192 // The merged store can be at most 8 bytes. 3193 const uint merge_list_max_size = 8 / _store->memory_size(); 3194 assert(merge_list_max_size >= 2 && 3195 merge_list_max_size <= 8 && 3196 is_power_of_2(merge_list_max_size), 3197 "must be 2, 4 or 8"); 3198 3199 // Traverse up the chain of adjacent def stores. 3200 StoreNode* current = _store; 3201 merge_list.push(current); 3202 while (current != nullptr && merge_list.size() < merge_list_max_size) { 3203 Status status = find_adjacent_def_store(current); 3204 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] find def: "); status.print_on(tty); }) 3205 3206 current = status.found_store(); 3207 if (current != nullptr) { 3208 merge_list.push(current); 3209 3210 // We can have at most one RangeCheck. 3211 if (status.found_range_check()) { 3212 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found RangeCheck, stop traversal."); }) 3213 break; 3214 } 3215 } 3216 } 3217 3218 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found:"); merge_list.dump(); }) 3219 3220 // Truncate the merge_list to a power of 2. 3221 const uint pow2size = round_down_power_of_2(merge_list.size()); 3222 assert(pow2size >= 2, "must be merging at least 2 stores"); 3223 while (merge_list.size() > pow2size) { merge_list.pop(); } 3224 3225 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] truncated:"); merge_list.dump(); }) 3226 } 3227 3228 // Merge the input values of the smaller stores to a single larger input value. 3229 Node* MergePrimitiveStores::make_merged_input_value(const Node_List& merge_list) { 3230 int new_memory_size = _store->memory_size() * merge_list.size(); 3231 Node* first = merge_list.at(merge_list.size()-1); 3232 Node* merged_input_value = nullptr; 3233 if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) { 3234 assert(_value_order == ValueOrder::Const, "must match"); 3235 // Pattern: [ConI, ConI, ...] -> new constant 3236 jlong con = 0; 3237 jlong bits_per_store = _store->memory_size() * 8; 3238 jlong mask = (((jlong)1) << bits_per_store) - 1; 3239 for (uint i = 0; i < merge_list.size(); i++) { 3240 jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int(); 3241 #ifdef VM_LITTLE_ENDIAN 3242 con = con << bits_per_store; 3243 con = con | (mask & con_i); 3244 #else // VM_LITTLE_ENDIAN 3245 con_i = (mask & con_i) << (i * bits_per_store); 3246 con = con | con_i; 3247 #endif // VM_LITTLE_ENDIAN 3248 } 3249 merged_input_value = _phase->longcon(con); 3250 } else { 3251 assert(_value_order == ValueOrder::Platform || _value_order == ValueOrder::Reverse, "must match"); 3252 // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base 3253 // | | 3254 // _store first 3255 // 3256 Node* hi = _store->in(MemNode::ValueIn); 3257 Node* lo = first->in(MemNode::ValueIn); 3258 #ifndef VM_LITTLE_ENDIAN 3259 // `_store` and `first` are swapped in the diagram above 3260 swap(hi, lo); 3261 #endif // !VM_LITTLE_ENDIAN 3262 if (_value_order == ValueOrder::Reverse) { 3263 swap(hi, lo); 3264 } 3265 Node const* hi_base; 3266 jint hi_shift; 3267 merged_input_value = lo; 3268 bool is_true = is_con_RShift(hi, hi_base, hi_shift, _phase); 3269 assert(is_true, "must detect con RShift"); 3270 if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) { 3271 // look through 3272 merged_input_value = merged_input_value->in(1); 3273 } 3274 if (merged_input_value != hi_base) { 3275 // merged_input_value is not the base 3276 return nullptr; 3277 } 3278 } 3279 3280 if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) { 3281 // Example: 3282 // 3283 // long base = ...; 3284 // a[0] = (byte)(base >> 0); 3285 // a[1] = (byte)(base >> 8); 3286 // 3287 merged_input_value = _phase->transform(new ConvL2INode(merged_input_value)); 3288 } 3289 3290 assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) || 3291 (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8), 3292 "merged_input_value is either int or long, and new_memory_size is small enough"); 3293 3294 if (_value_order == ValueOrder::Reverse) { 3295 assert(_store->memory_size() == 1, "only implemented for bytes"); 3296 if (new_memory_size == 8) { 3297 merged_input_value = _phase->transform(new ReverseBytesLNode(merged_input_value)); 3298 } else if (new_memory_size == 4) { 3299 merged_input_value = _phase->transform(new ReverseBytesINode(merged_input_value)); 3300 } else { 3301 assert(new_memory_size == 2, "sanity check"); 3302 merged_input_value = _phase->transform(new ReverseBytesSNode(merged_input_value)); 3303 } 3304 } 3305 return merged_input_value; 3306 } 3307 3308 // // 3309 // first_ctrl first_mem first_adr first_ctrl first_mem first_adr // 3310 // | | | | | | // 3311 // | | | | +---------------+ | // 3312 // | | | | | | | // 3313 // | | +---------+ | | +---------------+ // 3314 // | | | | | | | | // 3315 // +--------------+ | | v1 +------------------------------+ | | v1 // 3316 // | | | | | | | | | | | | // 3317 // RangeCheck first_store RangeCheck | | first_store // 3318 // | | | | | | | // 3319 // last_ctrl | +----> unc_trap last_ctrl | | +----> unc_trap // 3320 // | | ===> | | | // 3321 // +--------------+ | a2 v2 | | | // 3322 // | | | | | | | | // 3323 // | second_store | | | // 3324 // | | | | | [v1 v2 ... vn] // 3325 // ... ... | | | | // 3326 // | | | | | v // 3327 // +--------------+ | an vn +--------------+ | | merged_input_value // 3328 // | | | | | | | | // 3329 // last_store (= _store) merged_store // 3330 // // 3331 StoreNode* MergePrimitiveStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) { 3332 Node* first_store = merge_list.at(merge_list.size()-1); 3333 Node* last_ctrl = _store->in(MemNode::Control); // after (optional) RangeCheck 3334 Node* first_mem = first_store->in(MemNode::Memory); 3335 Node* first_adr = first_store->in(MemNode::Address); 3336 3337 const TypePtr* new_adr_type = _store->adr_type(); 3338 3339 int new_memory_size = _store->memory_size() * merge_list.size(); 3340 BasicType bt = T_ILLEGAL; 3341 switch (new_memory_size) { 3342 case 2: bt = T_SHORT; break; 3343 case 4: bt = T_INT; break; 3344 case 8: bt = T_LONG; break; 3345 } 3346 3347 StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr, 3348 new_adr_type, merged_input_value, bt, MemNode::unordered); 3349 3350 // Marking the store mismatched is sufficient to prevent reordering, since array stores 3351 // are all on the same slice. Hence, we need no barriers. 3352 merged_store->set_mismatched_access(); 3353 3354 // Constants above may now also be be packed -> put candidate on worklist 3355 _phase->is_IterGVN()->_worklist.push(first_mem); 3356 3357 return merged_store; 3358 } 3359 3360 #ifndef PRODUCT 3361 void MergePrimitiveStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const { 3362 stringStream ss; 3363 ss.print_cr("[TraceMergeStores]: Replace"); 3364 for (int i = (int)merge_list.size() - 1; i >= 0; i--) { 3365 merge_list.at(i)->dump("\n", false, &ss); 3366 } 3367 ss.print_cr("[TraceMergeStores]: with"); 3368 merged_input_value->dump("\n", false, &ss); 3369 merged_store->dump("\n", false, &ss); 3370 tty->print("%s", ss.as_string()); 3371 } 3372 #endif 3373 3374 //------------------------------Ideal------------------------------------------ 3375 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 3376 // When a store immediately follows a relevant allocation/initialization, 3377 // try to capture it into the initialization, or hoist it above. 3378 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3379 Node* p = MemNode::Ideal_common(phase, can_reshape); 3380 if (p) return (p == NodeSentinel) ? nullptr : p; 3381 3382 Node* mem = in(MemNode::Memory); 3383 Node* address = in(MemNode::Address); 3384 Node* value = in(MemNode::ValueIn); 3385 // Back-to-back stores to same address? Fold em up. Generally 3386 // unsafe if I have intervening uses. 3387 { 3388 Node* st = mem; 3389 // If Store 'st' has more than one use, we cannot fold 'st' away. 3390 // For example, 'st' might be the final state at a conditional 3391 // return. Or, 'st' might be used by some node which is live at 3392 // the same time 'st' is live, which might be unschedulable. So, 3393 // require exactly ONE user until such time as we clone 'mem' for 3394 // each of 'mem's uses (thus making the exactly-1-user-rule hold 3395 // true). 3396 while (st->is_Store() && st->outcnt() == 1) { 3397 // Looking at a dead closed cycle of memory? 3398 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 3399 assert(Opcode() == st->Opcode() || 3400 st->Opcode() == Op_StoreVector || 3401 Opcode() == Op_StoreVector || 3402 st->Opcode() == Op_StoreVectorScatter || 3403 Opcode() == Op_StoreVectorScatter || 3404 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 3405 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 3406 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy 3407 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 3408 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 3409 3410 if (st->in(MemNode::Address)->eqv_uncast(address) && 3411 st->as_Store()->memory_size() <= this->memory_size()) { 3412 Node* use = st->raw_out(0); 3413 if (phase->is_IterGVN()) { 3414 phase->is_IterGVN()->rehash_node_delayed(use); 3415 } 3416 // It's OK to do this in the parser, since DU info is always accurate, 3417 // and the parser always refers to nodes via SafePointNode maps. 3418 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase); 3419 return this; 3420 } 3421 st = st->in(MemNode::Memory); 3422 } 3423 } 3424 3425 3426 // Capture an unaliased, unconditional, simple store into an initializer. 3427 // Or, if it is independent of the allocation, hoist it above the allocation. 3428 if (ReduceFieldZeroing && /*can_reshape &&*/ 3429 mem->is_Proj() && mem->in(0)->is_Initialize()) { 3430 InitializeNode* init = mem->in(0)->as_Initialize(); 3431 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 3432 if (offset > 0) { 3433 Node* moved = init->capture_store(this, offset, phase, can_reshape); 3434 // If the InitializeNode captured me, it made a raw copy of me, 3435 // and I need to disappear. 3436 if (moved != nullptr) { 3437 // %%% hack to ensure that Ideal returns a new node: 3438 mem = MergeMemNode::make(mem); 3439 return mem; // fold me away 3440 } 3441 } 3442 } 3443 3444 // Fold reinterpret cast into memory operation: 3445 // StoreX mem (MoveY2X v) => StoreY mem v 3446 if (value->is_Move()) { 3447 const Type* vt = value->in(1)->bottom_type(); 3448 if (has_reinterpret_variant(vt)) { 3449 if (phase->C->post_loop_opts_phase()) { 3450 return convert_to_reinterpret_store(*phase, value->in(1), vt); 3451 } else { 3452 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over 3453 } 3454 } 3455 } 3456 3457 if (MergeStores && UseUnalignedAccesses) { 3458 if (phase->C->merge_stores_phase()) { 3459 MergePrimitiveStores merge(phase, this); 3460 Node* progress = merge.run(); 3461 if (progress != nullptr) { return progress; } 3462 } else { 3463 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during 3464 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between 3465 // the stores, and we merge the wrong sequence of stores. 3466 // Example: 3467 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI 3468 // Apply MergeStores: 3469 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI 3470 // Remove more RangeChecks: 3471 // StoreI [ StoreL ] StoreI 3472 // But now it would have been better to do this instead: 3473 // [ StoreL ] [ StoreL ] 3474 phase->C->record_for_merge_stores_igvn(this); 3475 } 3476 } 3477 3478 return nullptr; // No further progress 3479 } 3480 3481 //------------------------------Value----------------------------------------- 3482 const Type* StoreNode::Value(PhaseGVN* phase) const { 3483 // Either input is TOP ==> the result is TOP 3484 const Type *t1 = phase->type( in(MemNode::Memory) ); 3485 if( t1 == Type::TOP ) return Type::TOP; 3486 const Type *t2 = phase->type( in(MemNode::Address) ); 3487 if( t2 == Type::TOP ) return Type::TOP; 3488 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 3489 if( t3 == Type::TOP ) return Type::TOP; 3490 return Type::MEMORY; 3491 } 3492 3493 //------------------------------Identity--------------------------------------- 3494 // Remove redundant stores: 3495 // Store(m, p, Load(m, p)) changes to m. 3496 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 3497 Node* StoreNode::Identity(PhaseGVN* phase) { 3498 Node* mem = in(MemNode::Memory); 3499 Node* adr = in(MemNode::Address); 3500 Node* val = in(MemNode::ValueIn); 3501 3502 Node* result = this; 3503 3504 // Load then Store? Then the Store is useless 3505 if (val->is_Load() && 3506 val->in(MemNode::Address)->eqv_uncast(adr) && 3507 val->in(MemNode::Memory )->eqv_uncast(mem) && 3508 val->as_Load()->store_Opcode() == Opcode()) { 3509 // Ensure vector type is the same 3510 if (!is_StoreVector() || (mem->is_LoadVector() && as_StoreVector()->vect_type() == mem->as_LoadVector()->vect_type())) { 3511 result = mem; 3512 } 3513 } 3514 3515 // Two stores in a row of the same value? 3516 if (result == this && 3517 mem->is_Store() && 3518 mem->in(MemNode::Address)->eqv_uncast(adr) && 3519 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 3520 mem->Opcode() == Opcode()) { 3521 if (!is_StoreVector()) { 3522 result = mem; 3523 } else { 3524 const StoreVectorNode* store_vector = as_StoreVector(); 3525 const StoreVectorNode* mem_vector = mem->as_StoreVector(); 3526 const Node* store_indices = store_vector->indices(); 3527 const Node* mem_indices = mem_vector->indices(); 3528 const Node* store_mask = store_vector->mask(); 3529 const Node* mem_mask = mem_vector->mask(); 3530 // Ensure types, indices, and masks match 3531 if (store_vector->vect_type() == mem_vector->vect_type() && 3532 ((store_indices == nullptr) == (mem_indices == nullptr) && 3533 (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) && 3534 ((store_mask == nullptr) == (mem_mask == nullptr) && 3535 (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) { 3536 result = mem; 3537 } 3538 } 3539 } 3540 3541 // Store of zero anywhere into a freshly-allocated object? 3542 // Then the store is useless. 3543 // (It must already have been captured by the InitializeNode.) 3544 if (result == this && 3545 ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 3546 // a newly allocated object is already all-zeroes everywhere 3547 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 3548 result = mem; 3549 } 3550 3551 if (result == this) { 3552 // the store may also apply to zero-bits in an earlier object 3553 Node* prev_mem = find_previous_store(phase); 3554 // Steps (a), (b): Walk past independent stores to find an exact match. 3555 if (prev_mem != nullptr) { 3556 Node* prev_val = can_see_stored_value(prev_mem, phase); 3557 if (prev_val != nullptr && prev_val == val) { 3558 // prev_val and val might differ by a cast; it would be good 3559 // to keep the more informative of the two. 3560 result = mem; 3561 } 3562 } 3563 } 3564 } 3565 3566 PhaseIterGVN* igvn = phase->is_IterGVN(); 3567 if (result != this && igvn != nullptr) { 3568 MemBarNode* trailing = trailing_membar(); 3569 if (trailing != nullptr) { 3570 #ifdef ASSERT 3571 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr(); 3572 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects"); 3573 #endif 3574 trailing->remove(igvn); 3575 } 3576 } 3577 3578 return result; 3579 } 3580 3581 //------------------------------match_edge------------------------------------- 3582 // Do we Match on this edge index or not? Match only memory & value 3583 uint StoreNode::match_edge(uint idx) const { 3584 return idx == MemNode::Address || idx == MemNode::ValueIn; 3585 } 3586 3587 //------------------------------cmp-------------------------------------------- 3588 // Do not common stores up together. They generally have to be split 3589 // back up anyways, so do not bother. 3590 bool StoreNode::cmp( const Node &n ) const { 3591 return (&n == this); // Always fail except on self 3592 } 3593 3594 //------------------------------Ideal_masked_input----------------------------- 3595 // Check for a useless mask before a partial-word store 3596 // (StoreB ... (AndI valIn conIa) ) 3597 // If (conIa & mask == mask) this simplifies to 3598 // (StoreB ... (valIn) ) 3599 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 3600 Node *val = in(MemNode::ValueIn); 3601 if( val->Opcode() == Op_AndI ) { 3602 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 3603 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 3604 set_req_X(MemNode::ValueIn, val->in(1), phase); 3605 return this; 3606 } 3607 } 3608 return nullptr; 3609 } 3610 3611 3612 //------------------------------Ideal_sign_extended_input---------------------- 3613 // Check for useless sign-extension before a partial-word store 3614 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR)) 3615 // If (conIL == conIR && conIR <= num_rejected_bits) this simplifies to 3616 // (StoreB ... (v)) 3617 // If (conIL > conIR) under some conditions, it can be simplified into 3618 // (StoreB ... (LShiftI _ v (conIL - conIR))) 3619 // This case happens when the value of the store was itself a left shift, that 3620 // gets merged into the inner left shift of the sign-extension. For instance, 3621 // if we have 3622 // array_of_shorts[0] = (short)(v << 2) 3623 // We get a structure such as: 3624 // (StoreB ... (RShiftI _ (LShiftI _ (LShiftI _ v 2) 16) 16)) 3625 // that is simplified into 3626 // (StoreB ... (RShiftI _ (LShiftI _ v 18) 16)). 3627 // It is thus useful to handle cases where conIL > conIR. But this simplification 3628 // does not always hold. Let's see in which cases it's valid. 3629 // 3630 // Let's assume we have the following 32 bits integer v: 3631 // +----------------------------------+ 3632 // | v[0..31] | 3633 // +----------------------------------+ 3634 // 31 0 3635 // that will be stuffed in 8 bits byte after a shift left and a shift right of 3636 // potentially different magnitudes. 3637 // We denote num_rejected_bits the number of bits of the discarded part. In this 3638 // case, num_rejected_bits == 24. 3639 // 3640 // Statement (proved further below in case analysis): 3641 // Given: 3642 // - 0 <= conIL < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount) 3643 // - 0 <= conIR < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount) 3644 // - conIL >= conIR 3645 // - num_rejected_bits >= conIR 3646 // Then this form: 3647 // (RShiftI _ (LShiftI _ v conIL) conIR) 3648 // can be replaced with this form: 3649 // (LShiftI _ v (conIL-conIR)) 3650 // 3651 // Note: We only have to show that the non-rejected lowest bits (8 bits for byte) 3652 // have to be correct, as the higher bits are rejected / truncated by the store. 3653 // 3654 // The hypotheses 3655 // 0 <= conIL < BitsPerJavaInteger 3656 // 0 <= conIR < BitsPerJavaInteger 3657 // are ensured by maskShiftAmount (called from ::Ideal of shift nodes). Indeed, 3658 // (v << 31) << 2 must be simplified into 0, not into v << 33 (which is equivalent 3659 // to v << 1). 3660 // 3661 // 3662 // If you don't like case analysis, jump after the conclusion. 3663 // ### Case 1 : conIL == conIR 3664 // ###### Case 1.1: conIL == conIR == num_rejected_bits 3665 // If we do the shift left then right by 24 bits, we get: 3666 // after: v << 24 3667 // +---------+------------------------+ 3668 // | v[0..7] | 0 | 3669 // +---------+------------------------+ 3670 // 31 24 23 0 3671 // after: (v << 24) >> 24 3672 // +------------------------+---------+ 3673 // | sign bit | v[0..7] | 3674 // +------------------------+---------+ 3675 // 31 8 7 0 3676 // The non-rejected bits (bits kept by the store, that is the 8 lower bits of the 3677 // result) are the same before and after, so, indeed, simplifying is correct. 3678 3679 // ###### Case 1.2: conIL == conIR < num_rejected_bits 3680 // If we do the shift left then right by 22 bits, we get: 3681 // after: v << 22 3682 // +---------+------------------------+ 3683 // | v[0..9] | 0 | 3684 // +---------+------------------------+ 3685 // 31 22 21 0 3686 // after: (v << 22) >> 22 3687 // +------------------------+---------+ 3688 // | sign bit | v[0..9] | 3689 // +------------------------+---------+ 3690 // 31 10 9 0 3691 // The non-rejected bits are the 8 lower bits of v. The bits 8 and 9 of v are still 3692 // present in (v << 22) >> 22 but will be dropped by the store. The simplification is 3693 // still correct. 3694 3695 // ###### But! Case 1.3: conIL == conIR > num_rejected_bits 3696 // If we do the shift left then right by 26 bits, we get: 3697 // after: v << 26 3698 // +---------+------------------------+ 3699 // | v[0..5] | 0 | 3700 // +---------+------------------------+ 3701 // 31 26 25 0 3702 // after: (v << 26) >> 26 3703 // +------------------------+---------+ 3704 // | sign bit | v[0..5] | 3705 // +------------------------+---------+ 3706 // 31 6 5 0 3707 // The non-rejected bits are made of 3708 // - 0-5 => the bits 0 to 5 of v 3709 // - 6-7 => the sign bit of v[0..5] (that is v[5]) 3710 // Simplifying this as v is not correct. 3711 // The condition conIR <= num_rejected_bits is indeed necessary in Case 1 3712 // 3713 // ### Case 2: conIL > conIR 3714 // ###### Case 2.1: num_rejected_bits == conIR 3715 // We take conIL == 26 for this example. 3716 // after: v << 26 3717 // +---------+------------------------+ 3718 // | v[0..5] | 0 | 3719 // +---------+------------------------+ 3720 // 31 26 25 0 3721 // after: (v << 26) >> 24 3722 // +------------------+---------+-----+ 3723 // | sign bit | v[0..5] | 0 | 3724 // +------------------+---------+-----+ 3725 // 31 8 7 2 1 0 3726 // The non-rejected bits are the 8 lower ones of (v << conIL - conIR). 3727 // The bits 6 and 7 of v have been thrown away after the shift left. 3728 // The simplification is still correct. 3729 // 3730 // ###### Case 2.2: num_rejected_bits > conIR. 3731 // Let's say conIL == 26 and conIR == 22. 3732 // after: v << 26 3733 // +---------+------------------------+ 3734 // | v[0..5] | 0 | 3735 // +---------+------------------------+ 3736 // 31 26 25 0 3737 // after: (v << 26) >> 22 3738 // +------------------+---------+-----+ 3739 // | sign bit | v[0..5] | 0 | 3740 // +------------------+---------+-----+ 3741 // 31 10 9 4 3 0 3742 // The bits non-rejected by the store are exactly the 8 lower ones of (v << (conIL - conIR)): 3743 // - 0-3 => 0 3744 // - 4-7 => bits 0 to 3 of v 3745 // The simplification is still correct. 3746 // The bits 4 and 5 of v are still present in (v << (conIL - conIR)) but they don't 3747 // matter as they are not in the 8 lower bits: they will be cut out by the store. 3748 // 3749 // ###### But! Case 2.3: num_rejected_bits < conIR. 3750 // Let's see that this case is not as easy to simplify. 3751 // Let's say conIL == 28 and conIR == 26. 3752 // after: v << 28 3753 // +---------+------------------------+ 3754 // | v[0..3] | 0 | 3755 // +---------+------------------------+ 3756 // 31 28 27 0 3757 // after: (v << 28) >> 26 3758 // +------------------+---------+-----+ 3759 // | sign bit | v[0..3] | 0 | 3760 // +------------------+---------+-----+ 3761 // 31 6 5 2 1 0 3762 // The non-rejected bits are made of 3763 // - 0-1 => 0 3764 // - 2-5 => the bits 0 to 3 of v 3765 // - 6-7 => the sign bit of v[0..3] (that is v[3]) 3766 // Simplifying this as (v << 2) is not correct. 3767 // The condition conIR <= num_rejected_bits is indeed necessary in Case 2. 3768 // 3769 // ### Conclusion: 3770 // Our hypotheses are indeed sufficient: 3771 // - 0 <= conIL < BitsPerJavaInteger 3772 // - 0 <= conIR < BitsPerJavaInteger 3773 // - conIL >= conIR 3774 // - num_rejected_bits >= conIR 3775 // 3776 // ### A rationale without case analysis: 3777 // After the shift left, conIL upper bits of v are discarded and conIL lower bit 3778 // zeroes are added. After the shift right, conIR lower bits of the previous result 3779 // are discarded. If conIL >= conIR, we discard only the zeroes we made up during 3780 // the shift left, but if conIL < conIR, then we discard also lower bits of v. But 3781 // the point of the simplification is to get an expression of the form 3782 // (v << (conIL - conIR)). This expression discard only higher bits of v, thus the 3783 // simplification is not correct if conIL < conIR. 3784 // 3785 // Moreover, after the shift right, the higher bit of (v << conIL) is repeated on the 3786 // conIR higher bits of ((v << conIL) >> conIR), it's the sign-extension. If 3787 // conIR > num_rejected_bits, then at least one artificial copy of this sign bit will 3788 // be in the window of the store. Thus ((v << conIL) >> conIR) is not equivalent to 3789 // (v << (conIL-conIR)) if conIR > num_rejected_bits. 3790 // 3791 // We do not treat the case conIL < conIR here since the point of this function is 3792 // to skip sign-extensions (that is conIL == conIR == num_rejected_bits). The need 3793 // of treating conIL > conIR comes from the cases where the sign-extended value is 3794 // also left-shift expression. Computing the sign-extension of a right-shift expression 3795 // doesn't yield a situation such as 3796 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR)) 3797 // where conIL < conIR. 3798 Node* StoreNode::Ideal_sign_extended_input(PhaseGVN* phase, int num_rejected_bits) { 3799 Node* shr = in(MemNode::ValueIn); 3800 if (shr->Opcode() == Op_RShiftI) { 3801 const TypeInt* conIR = phase->type(shr->in(2))->isa_int(); 3802 if (conIR != nullptr && conIR->is_con() && conIR->get_con() >= 0 && conIR->get_con() < BitsPerJavaInteger && conIR->get_con() <= num_rejected_bits) { 3803 Node* shl = shr->in(1); 3804 if (shl->Opcode() == Op_LShiftI) { 3805 const TypeInt* conIL = phase->type(shl->in(2))->isa_int(); 3806 if (conIL != nullptr && conIL->is_con() && conIL->get_con() >= 0 && conIL->get_con() < BitsPerJavaInteger) { 3807 if (conIL->get_con() == conIR->get_con()) { 3808 set_req_X(MemNode::ValueIn, shl->in(1), phase); 3809 return this; 3810 } 3811 if (conIL->get_con() > conIR->get_con()) { 3812 Node* new_shl = phase->transform(new LShiftINode(shl->in(1), phase->intcon(conIL->get_con() - conIR->get_con()))); 3813 set_req_X(MemNode::ValueIn, new_shl, phase); 3814 return this; 3815 } 3816 } 3817 } 3818 } 3819 } 3820 return nullptr; 3821 } 3822 3823 //------------------------------value_never_loaded----------------------------------- 3824 // Determine whether there are any possible loads of the value stored. 3825 // For simplicity, we actually check if there are any loads from the 3826 // address stored to, not just for loads of the value stored by this node. 3827 // 3828 bool StoreNode::value_never_loaded(PhaseValues* phase) const { 3829 Node *adr = in(Address); 3830 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 3831 if (adr_oop == nullptr) 3832 return false; 3833 if (!adr_oop->is_known_instance_field()) 3834 return false; // if not a distinct instance, there may be aliases of the address 3835 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 3836 Node *use = adr->fast_out(i); 3837 if (use->is_Load() || use->is_LoadStore()) { 3838 return false; 3839 } 3840 } 3841 return true; 3842 } 3843 3844 MemBarNode* StoreNode::trailing_membar() const { 3845 if (is_release()) { 3846 MemBarNode* trailing_mb = nullptr; 3847 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 3848 Node* u = fast_out(i); 3849 if (u->is_MemBar()) { 3850 if (u->as_MemBar()->trailing_store()) { 3851 assert(u->Opcode() == Op_MemBarVolatile, ""); 3852 assert(trailing_mb == nullptr, "only one"); 3853 trailing_mb = u->as_MemBar(); 3854 #ifdef ASSERT 3855 Node* leading = u->as_MemBar()->leading_membar(); 3856 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 3857 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair"); 3858 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair"); 3859 #endif 3860 } else { 3861 assert(u->as_MemBar()->standalone(), ""); 3862 } 3863 } 3864 } 3865 return trailing_mb; 3866 } 3867 return nullptr; 3868 } 3869 3870 3871 //============================================================================= 3872 //------------------------------Ideal------------------------------------------ 3873 // If the store is from an AND mask that leaves the low bits untouched, then 3874 // we can skip the AND operation. If the store is from a sign-extension 3875 // (a left shift, then right shift) we can skip both. 3876 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 3877 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 3878 if( progress != nullptr ) return progress; 3879 3880 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 3881 if( progress != nullptr ) return progress; 3882 3883 // Finally check the default case 3884 return StoreNode::Ideal(phase, can_reshape); 3885 } 3886 3887 //============================================================================= 3888 //------------------------------Ideal------------------------------------------ 3889 // If the store is from an AND mask that leaves the low bits untouched, then 3890 // we can skip the AND operation 3891 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 3892 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 3893 if( progress != nullptr ) return progress; 3894 3895 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 3896 if( progress != nullptr ) return progress; 3897 3898 // Finally check the default case 3899 return StoreNode::Ideal(phase, can_reshape); 3900 } 3901 3902 //============================================================================= 3903 //----------------------------------SCMemProjNode------------------------------ 3904 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 3905 { 3906 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) { 3907 return Type::TOP; 3908 } 3909 return bottom_type(); 3910 } 3911 3912 //============================================================================= 3913 //----------------------------------LoadStoreNode------------------------------ 3914 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 3915 : Node(required), 3916 _type(rt), 3917 _adr_type(at), 3918 _barrier_data(0) 3919 { 3920 init_req(MemNode::Control, c ); 3921 init_req(MemNode::Memory , mem); 3922 init_req(MemNode::Address, adr); 3923 init_req(MemNode::ValueIn, val); 3924 init_class_id(Class_LoadStore); 3925 } 3926 3927 //------------------------------Value----------------------------------------- 3928 const Type* LoadStoreNode::Value(PhaseGVN* phase) const { 3929 // Either input is TOP ==> the result is TOP 3930 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) { 3931 return Type::TOP; 3932 } 3933 const Type* t = phase->type(in(MemNode::Memory)); 3934 if (t == Type::TOP) { 3935 return Type::TOP; 3936 } 3937 t = phase->type(in(MemNode::Address)); 3938 if (t == Type::TOP) { 3939 return Type::TOP; 3940 } 3941 t = phase->type(in(MemNode::ValueIn)); 3942 if (t == Type::TOP) { 3943 return Type::TOP; 3944 } 3945 return bottom_type(); 3946 } 3947 3948 uint LoadStoreNode::ideal_reg() const { 3949 return _type->ideal_reg(); 3950 } 3951 3952 // This method conservatively checks if the result of a LoadStoreNode is 3953 // used, that is, if it returns true, then it is definitely the case that 3954 // the result of the node is not needed. 3955 // For example, GetAndAdd can be matched into a lock_add instead of a 3956 // lock_xadd if the result of LoadStoreNode::result_not_used() is true 3957 bool LoadStoreNode::result_not_used() const { 3958 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 3959 Node *x = fast_out(i); 3960 if (x->Opcode() == Op_SCMemProj) { 3961 continue; 3962 } 3963 if (x->bottom_type() == TypeTuple::MEMBAR && 3964 !x->is_Call() && 3965 x->Opcode() != Op_Blackhole) { 3966 continue; 3967 } 3968 return false; 3969 } 3970 return true; 3971 } 3972 3973 MemBarNode* LoadStoreNode::trailing_membar() const { 3974 MemBarNode* trailing = nullptr; 3975 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 3976 Node* u = fast_out(i); 3977 if (u->is_MemBar()) { 3978 if (u->as_MemBar()->trailing_load_store()) { 3979 assert(u->Opcode() == Op_MemBarAcquire, ""); 3980 assert(trailing == nullptr, "only one"); 3981 trailing = u->as_MemBar(); 3982 #ifdef ASSERT 3983 Node* leading = trailing->leading_membar(); 3984 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 3985 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair"); 3986 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair"); 3987 #endif 3988 } else { 3989 assert(u->as_MemBar()->standalone(), "wrong barrier kind"); 3990 } 3991 } 3992 } 3993 3994 return trailing; 3995 } 3996 3997 uint LoadStoreNode::size_of() const { return sizeof(*this); } 3998 3999 //============================================================================= 4000 //----------------------------------LoadStoreConditionalNode-------------------- 4001 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) { 4002 init_req(ExpectedIn, ex ); 4003 } 4004 4005 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const { 4006 // Either input is TOP ==> the result is TOP 4007 const Type* t = phase->type(in(ExpectedIn)); 4008 if (t == Type::TOP) { 4009 return Type::TOP; 4010 } 4011 return LoadStoreNode::Value(phase); 4012 } 4013 4014 //============================================================================= 4015 //-------------------------------adr_type-------------------------------------- 4016 const TypePtr* ClearArrayNode::adr_type() const { 4017 Node *adr = in(3); 4018 if (adr == nullptr) return nullptr; // node is dead 4019 return MemNode::calculate_adr_type(adr->bottom_type()); 4020 } 4021 4022 //------------------------------match_edge------------------------------------- 4023 // Do we Match on this edge index or not? Do not match memory 4024 uint ClearArrayNode::match_edge(uint idx) const { 4025 return idx > 1; 4026 } 4027 4028 //------------------------------Identity--------------------------------------- 4029 // Clearing a zero length array does nothing 4030 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 4031 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 4032 } 4033 4034 //------------------------------Idealize--------------------------------------- 4035 // Clearing a short array is faster with stores 4036 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4037 // Already know this is a large node, do not try to ideal it 4038 if (_is_large) return nullptr; 4039 4040 const int unit = BytesPerLong; 4041 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 4042 if (!t) return nullptr; 4043 if (!t->is_con()) return nullptr; 4044 intptr_t raw_count = t->get_con(); 4045 intptr_t size = raw_count; 4046 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 4047 // Clearing nothing uses the Identity call. 4048 // Negative clears are possible on dead ClearArrays 4049 // (see jck test stmt114.stmt11402.val). 4050 if (size <= 0 || size % unit != 0) return nullptr; 4051 intptr_t count = size / unit; 4052 // Length too long; communicate this to matchers and assemblers. 4053 // Assemblers are responsible to produce fast hardware clears for it. 4054 if (size > InitArrayShortSize) { 4055 return new ClearArrayNode(in(0), in(1), in(2), in(3), true); 4056 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) { 4057 return nullptr; 4058 } 4059 if (!IdealizeClearArrayNode) return nullptr; 4060 Node *mem = in(1); 4061 if( phase->type(mem)==Type::TOP ) return nullptr; 4062 Node *adr = in(3); 4063 const Type* at = phase->type(adr); 4064 if( at==Type::TOP ) return nullptr; 4065 const TypePtr* atp = at->isa_ptr(); 4066 // adjust atp to be the correct array element address type 4067 if (atp == nullptr) atp = TypePtr::BOTTOM; 4068 else atp = atp->add_offset(Type::OffsetBot); 4069 // Get base for derived pointer purposes 4070 if( adr->Opcode() != Op_AddP ) Unimplemented(); 4071 Node *base = adr->in(1); 4072 4073 Node *zero = phase->makecon(TypeLong::ZERO); 4074 Node *off = phase->MakeConX(BytesPerLong); 4075 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 4076 count--; 4077 while( count-- ) { 4078 mem = phase->transform(mem); 4079 adr = phase->transform(new AddPNode(base,adr,off)); 4080 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 4081 } 4082 return mem; 4083 } 4084 4085 //----------------------------step_through---------------------------------- 4086 // Return allocation input memory edge if it is different instance 4087 // or itself if it is the one we are looking for. 4088 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) { 4089 Node* n = *np; 4090 assert(n->is_ClearArray(), "sanity"); 4091 intptr_t offset; 4092 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 4093 // This method is called only before Allocate nodes are expanded 4094 // during macro nodes expansion. Before that ClearArray nodes are 4095 // only generated in PhaseMacroExpand::generate_arraycopy() (before 4096 // Allocate nodes are expanded) which follows allocations. 4097 assert(alloc != nullptr, "should have allocation"); 4098 if (alloc->_idx == instance_id) { 4099 // Can not bypass initialization of the instance we are looking for. 4100 return false; 4101 } 4102 // Otherwise skip it. 4103 InitializeNode* init = alloc->initialization(); 4104 if (init != nullptr) 4105 *np = init->in(TypeFunc::Memory); 4106 else 4107 *np = alloc->in(TypeFunc::Memory); 4108 return true; 4109 } 4110 4111 //----------------------------clear_memory------------------------------------- 4112 // Generate code to initialize object storage to zero. 4113 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 4114 intptr_t start_offset, 4115 Node* end_offset, 4116 PhaseGVN* phase) { 4117 intptr_t offset = start_offset; 4118 4119 int unit = BytesPerLong; 4120 if ((offset % unit) != 0) { 4121 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 4122 adr = phase->transform(adr); 4123 const TypePtr* atp = TypeRawPtr::BOTTOM; 4124 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 4125 mem = phase->transform(mem); 4126 offset += BytesPerInt; 4127 } 4128 assert((offset % unit) == 0, ""); 4129 4130 // Initialize the remaining stuff, if any, with a ClearArray. 4131 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 4132 } 4133 4134 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 4135 Node* start_offset, 4136 Node* end_offset, 4137 PhaseGVN* phase) { 4138 if (start_offset == end_offset) { 4139 // nothing to do 4140 return mem; 4141 } 4142 4143 int unit = BytesPerLong; 4144 Node* zbase = start_offset; 4145 Node* zend = end_offset; 4146 4147 // Scale to the unit required by the CPU: 4148 if (!Matcher::init_array_count_is_in_bytes) { 4149 Node* shift = phase->intcon(exact_log2(unit)); 4150 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 4151 zend = phase->transform(new URShiftXNode(zend, shift) ); 4152 } 4153 4154 // Bulk clear double-words 4155 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 4156 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 4157 mem = new ClearArrayNode(ctl, mem, zsize, adr, false); 4158 return phase->transform(mem); 4159 } 4160 4161 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 4162 intptr_t start_offset, 4163 intptr_t end_offset, 4164 PhaseGVN* phase) { 4165 if (start_offset == end_offset) { 4166 // nothing to do 4167 return mem; 4168 } 4169 4170 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 4171 intptr_t done_offset = end_offset; 4172 if ((done_offset % BytesPerLong) != 0) { 4173 done_offset -= BytesPerInt; 4174 } 4175 if (done_offset > start_offset) { 4176 mem = clear_memory(ctl, mem, dest, 4177 start_offset, phase->MakeConX(done_offset), phase); 4178 } 4179 if (done_offset < end_offset) { // emit the final 32-bit store 4180 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 4181 adr = phase->transform(adr); 4182 const TypePtr* atp = TypeRawPtr::BOTTOM; 4183 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 4184 mem = phase->transform(mem); 4185 done_offset += BytesPerInt; 4186 } 4187 assert(done_offset == end_offset, ""); 4188 return mem; 4189 } 4190 4191 //============================================================================= 4192 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 4193 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)), 4194 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone) 4195 #ifdef ASSERT 4196 , _pair_idx(0) 4197 #endif 4198 { 4199 init_class_id(Class_MemBar); 4200 Node* top = C->top(); 4201 init_req(TypeFunc::I_O,top); 4202 init_req(TypeFunc::FramePtr,top); 4203 init_req(TypeFunc::ReturnAdr,top); 4204 if (precedent != nullptr) 4205 init_req(TypeFunc::Parms, precedent); 4206 } 4207 4208 //------------------------------cmp-------------------------------------------- 4209 uint MemBarNode::hash() const { return NO_HASH; } 4210 bool MemBarNode::cmp( const Node &n ) const { 4211 return (&n == this); // Always fail except on self 4212 } 4213 4214 //------------------------------make------------------------------------------- 4215 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 4216 switch (opcode) { 4217 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 4218 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 4219 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 4220 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 4221 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 4222 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn); 4223 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 4224 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 4225 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 4226 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 4227 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 4228 case Op_Initialize: return new InitializeNode(C, atp, pn); 4229 default: ShouldNotReachHere(); return nullptr; 4230 } 4231 } 4232 4233 void MemBarNode::remove(PhaseIterGVN *igvn) { 4234 if (outcnt() != 2) { 4235 assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory"); 4236 assert(outcnt() == 1, "Only control then"); 4237 } 4238 if (trailing_store() || trailing_load_store()) { 4239 MemBarNode* leading = leading_membar(); 4240 if (leading != nullptr) { 4241 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars"); 4242 leading->remove(igvn); 4243 } 4244 } 4245 if (proj_out_or_null(TypeFunc::Memory) != nullptr) { 4246 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 4247 } 4248 if (proj_out_or_null(TypeFunc::Control) != nullptr) { 4249 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 4250 } 4251 } 4252 4253 //------------------------------Ideal------------------------------------------ 4254 // Return a node which is more "ideal" than the current node. Strip out 4255 // control copies 4256 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4257 if (remove_dead_region(phase, can_reshape)) return this; 4258 // Don't bother trying to transform a dead node 4259 if (in(0) && in(0)->is_top()) { 4260 return nullptr; 4261 } 4262 4263 bool progress = false; 4264 // Eliminate volatile MemBars for scalar replaced objects. 4265 if (can_reshape && req() == (Precedent+1)) { 4266 bool eliminate = false; 4267 int opc = Opcode(); 4268 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 4269 // Volatile field loads and stores. 4270 Node* my_mem = in(MemBarNode::Precedent); 4271 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 4272 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 4273 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 4274 // replace this Precedent (decodeN) with the Load instead. 4275 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 4276 Node* load_node = my_mem->in(1); 4277 set_req(MemBarNode::Precedent, load_node); 4278 phase->is_IterGVN()->_worklist.push(my_mem); 4279 my_mem = load_node; 4280 } else { 4281 assert(my_mem->unique_out() == this, "sanity"); 4282 assert(!trailing_load_store(), "load store node can't be eliminated"); 4283 del_req(Precedent); 4284 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 4285 my_mem = nullptr; 4286 } 4287 progress = true; 4288 } 4289 if (my_mem != nullptr && my_mem->is_Mem()) { 4290 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 4291 // Check for scalar replaced object reference. 4292 if( t_oop != nullptr && t_oop->is_known_instance_field() && 4293 t_oop->offset() != Type::OffsetBot && 4294 t_oop->offset() != Type::OffsetTop) { 4295 eliminate = true; 4296 } 4297 } 4298 } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) { 4299 // Final field stores. 4300 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent)); 4301 if ((alloc != nullptr) && alloc->is_Allocate() && 4302 alloc->as_Allocate()->does_not_escape_thread()) { 4303 // The allocated object does not escape. 4304 eliminate = true; 4305 } 4306 } 4307 if (eliminate) { 4308 // Replace MemBar projections by its inputs. 4309 PhaseIterGVN* igvn = phase->is_IterGVN(); 4310 remove(igvn); 4311 // Must return either the original node (now dead) or a new node 4312 // (Do not return a top here, since that would break the uniqueness of top.) 4313 return new ConINode(TypeInt::ZERO); 4314 } 4315 } 4316 return progress ? this : nullptr; 4317 } 4318 4319 //------------------------------Value------------------------------------------ 4320 const Type* MemBarNode::Value(PhaseGVN* phase) const { 4321 if( !in(0) ) return Type::TOP; 4322 if( phase->type(in(0)) == Type::TOP ) 4323 return Type::TOP; 4324 return TypeTuple::MEMBAR; 4325 } 4326 4327 //------------------------------match------------------------------------------ 4328 // Construct projections for memory. 4329 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 4330 switch (proj->_con) { 4331 case TypeFunc::Control: 4332 case TypeFunc::Memory: 4333 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 4334 } 4335 ShouldNotReachHere(); 4336 return nullptr; 4337 } 4338 4339 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) { 4340 trailing->_kind = TrailingStore; 4341 leading->_kind = LeadingStore; 4342 #ifdef ASSERT 4343 trailing->_pair_idx = leading->_idx; 4344 leading->_pair_idx = leading->_idx; 4345 #endif 4346 } 4347 4348 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) { 4349 trailing->_kind = TrailingLoadStore; 4350 leading->_kind = LeadingLoadStore; 4351 #ifdef ASSERT 4352 trailing->_pair_idx = leading->_idx; 4353 leading->_pair_idx = leading->_idx; 4354 #endif 4355 } 4356 4357 MemBarNode* MemBarNode::trailing_membar() const { 4358 ResourceMark rm; 4359 Node* trailing = (Node*)this; 4360 VectorSet seen; 4361 Node_Stack multis(0); 4362 do { 4363 Node* c = trailing; 4364 uint i = 0; 4365 do { 4366 trailing = nullptr; 4367 for (; i < c->outcnt(); i++) { 4368 Node* next = c->raw_out(i); 4369 if (next != c && next->is_CFG()) { 4370 if (c->is_MultiBranch()) { 4371 if (multis.node() == c) { 4372 multis.set_index(i+1); 4373 } else { 4374 multis.push(c, i+1); 4375 } 4376 } 4377 trailing = next; 4378 break; 4379 } 4380 } 4381 if (trailing != nullptr && !seen.test_set(trailing->_idx)) { 4382 break; 4383 } 4384 while (multis.size() > 0) { 4385 c = multis.node(); 4386 i = multis.index(); 4387 if (i < c->req()) { 4388 break; 4389 } 4390 multis.pop(); 4391 } 4392 } while (multis.size() > 0); 4393 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing()); 4394 4395 MemBarNode* mb = trailing->as_MemBar(); 4396 assert((mb->_kind == TrailingStore && _kind == LeadingStore) || 4397 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar"); 4398 assert(mb->_pair_idx == _pair_idx, "bad trailing membar"); 4399 return mb; 4400 } 4401 4402 MemBarNode* MemBarNode::leading_membar() const { 4403 ResourceMark rm; 4404 VectorSet seen; 4405 Node_Stack regions(0); 4406 Node* leading = in(0); 4407 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) { 4408 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) { 4409 leading = nullptr; 4410 while (regions.size() > 0 && leading == nullptr) { 4411 Node* r = regions.node(); 4412 uint i = regions.index(); 4413 if (i < r->req()) { 4414 leading = r->in(i); 4415 regions.set_index(i+1); 4416 } else { 4417 regions.pop(); 4418 } 4419 } 4420 if (leading == nullptr) { 4421 assert(regions.size() == 0, "all paths should have been tried"); 4422 return nullptr; 4423 } 4424 } 4425 if (leading->is_Region()) { 4426 regions.push(leading, 2); 4427 leading = leading->in(1); 4428 } else { 4429 leading = leading->in(0); 4430 } 4431 } 4432 #ifdef ASSERT 4433 Unique_Node_List wq; 4434 wq.push((Node*)this); 4435 uint found = 0; 4436 for (uint i = 0; i < wq.size(); i++) { 4437 Node* n = wq.at(i); 4438 if (n->is_Region()) { 4439 for (uint j = 1; j < n->req(); j++) { 4440 Node* in = n->in(j); 4441 if (in != nullptr && !in->is_top()) { 4442 wq.push(in); 4443 } 4444 } 4445 } else { 4446 if (n->is_MemBar() && n->as_MemBar()->leading()) { 4447 assert(n == leading, "consistency check failed"); 4448 found++; 4449 } else { 4450 Node* in = n->in(0); 4451 if (in != nullptr && !in->is_top()) { 4452 wq.push(in); 4453 } 4454 } 4455 } 4456 } 4457 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed"); 4458 #endif 4459 if (leading == nullptr) { 4460 return nullptr; 4461 } 4462 MemBarNode* mb = leading->as_MemBar(); 4463 assert((mb->_kind == LeadingStore && _kind == TrailingStore) || 4464 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar"); 4465 assert(mb->_pair_idx == _pair_idx, "bad leading membar"); 4466 return mb; 4467 } 4468 4469 4470 //===========================InitializeNode==================================== 4471 // SUMMARY: 4472 // This node acts as a memory barrier on raw memory, after some raw stores. 4473 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 4474 // The Initialize can 'capture' suitably constrained stores as raw inits. 4475 // It can coalesce related raw stores into larger units (called 'tiles'). 4476 // It can avoid zeroing new storage for memory units which have raw inits. 4477 // At macro-expansion, it is marked 'complete', and does not optimize further. 4478 // 4479 // EXAMPLE: 4480 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 4481 // ctl = incoming control; mem* = incoming memory 4482 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 4483 // First allocate uninitialized memory and fill in the header: 4484 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 4485 // ctl := alloc.Control; mem* := alloc.Memory* 4486 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 4487 // Then initialize to zero the non-header parts of the raw memory block: 4488 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 4489 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 4490 // After the initialize node executes, the object is ready for service: 4491 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 4492 // Suppose its body is immediately initialized as {1,2}: 4493 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 4494 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 4495 // mem.SLICE(#short[*]) := store2 4496 // 4497 // DETAILS: 4498 // An InitializeNode collects and isolates object initialization after 4499 // an AllocateNode and before the next possible safepoint. As a 4500 // memory barrier (MemBarNode), it keeps critical stores from drifting 4501 // down past any safepoint or any publication of the allocation. 4502 // Before this barrier, a newly-allocated object may have uninitialized bits. 4503 // After this barrier, it may be treated as a real oop, and GC is allowed. 4504 // 4505 // The semantics of the InitializeNode include an implicit zeroing of 4506 // the new object from object header to the end of the object. 4507 // (The object header and end are determined by the AllocateNode.) 4508 // 4509 // Certain stores may be added as direct inputs to the InitializeNode. 4510 // These stores must update raw memory, and they must be to addresses 4511 // derived from the raw address produced by AllocateNode, and with 4512 // a constant offset. They must be ordered by increasing offset. 4513 // The first one is at in(RawStores), the last at in(req()-1). 4514 // Unlike most memory operations, they are not linked in a chain, 4515 // but are displayed in parallel as users of the rawmem output of 4516 // the allocation. 4517 // 4518 // (See comments in InitializeNode::capture_store, which continue 4519 // the example given above.) 4520 // 4521 // When the associated Allocate is macro-expanded, the InitializeNode 4522 // may be rewritten to optimize collected stores. A ClearArrayNode 4523 // may also be created at that point to represent any required zeroing. 4524 // The InitializeNode is then marked 'complete', prohibiting further 4525 // capturing of nearby memory operations. 4526 // 4527 // During macro-expansion, all captured initializations which store 4528 // constant values of 32 bits or smaller are coalesced (if advantageous) 4529 // into larger 'tiles' 32 or 64 bits. This allows an object to be 4530 // initialized in fewer memory operations. Memory words which are 4531 // covered by neither tiles nor non-constant stores are pre-zeroed 4532 // by explicit stores of zero. (The code shape happens to do all 4533 // zeroing first, then all other stores, with both sequences occurring 4534 // in order of ascending offsets.) 4535 // 4536 // Alternatively, code may be inserted between an AllocateNode and its 4537 // InitializeNode, to perform arbitrary initialization of the new object. 4538 // E.g., the object copying intrinsics insert complex data transfers here. 4539 // The initialization must then be marked as 'complete' disable the 4540 // built-in zeroing semantics and the collection of initializing stores. 4541 // 4542 // While an InitializeNode is incomplete, reads from the memory state 4543 // produced by it are optimizable if they match the control edge and 4544 // new oop address associated with the allocation/initialization. 4545 // They return a stored value (if the offset matches) or else zero. 4546 // A write to the memory state, if it matches control and address, 4547 // and if it is to a constant offset, may be 'captured' by the 4548 // InitializeNode. It is cloned as a raw memory operation and rewired 4549 // inside the initialization, to the raw oop produced by the allocation. 4550 // Operations on addresses which are provably distinct (e.g., to 4551 // other AllocateNodes) are allowed to bypass the initialization. 4552 // 4553 // The effect of all this is to consolidate object initialization 4554 // (both arrays and non-arrays, both piecewise and bulk) into a 4555 // single location, where it can be optimized as a unit. 4556 // 4557 // Only stores with an offset less than TrackedInitializationLimit words 4558 // will be considered for capture by an InitializeNode. This puts a 4559 // reasonable limit on the complexity of optimized initializations. 4560 4561 //---------------------------InitializeNode------------------------------------ 4562 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 4563 : MemBarNode(C, adr_type, rawoop), 4564 _is_complete(Incomplete), _does_not_escape(false) 4565 { 4566 init_class_id(Class_Initialize); 4567 4568 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 4569 assert(in(RawAddress) == rawoop, "proper init"); 4570 // Note: allocation() can be null, for secondary initialization barriers 4571 } 4572 4573 // Since this node is not matched, it will be processed by the 4574 // register allocator. Declare that there are no constraints 4575 // on the allocation of the RawAddress edge. 4576 const RegMask &InitializeNode::in_RegMask(uint idx) const { 4577 // This edge should be set to top, by the set_complete. But be conservative. 4578 if (idx == InitializeNode::RawAddress) 4579 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 4580 return RegMask::Empty; 4581 } 4582 4583 Node* InitializeNode::memory(uint alias_idx) { 4584 Node* mem = in(Memory); 4585 if (mem->is_MergeMem()) { 4586 return mem->as_MergeMem()->memory_at(alias_idx); 4587 } else { 4588 // incoming raw memory is not split 4589 return mem; 4590 } 4591 } 4592 4593 bool InitializeNode::is_non_zero() { 4594 if (is_complete()) return false; 4595 remove_extra_zeroes(); 4596 return (req() > RawStores); 4597 } 4598 4599 void InitializeNode::set_complete(PhaseGVN* phase) { 4600 assert(!is_complete(), "caller responsibility"); 4601 _is_complete = Complete; 4602 4603 // After this node is complete, it contains a bunch of 4604 // raw-memory initializations. There is no need for 4605 // it to have anything to do with non-raw memory effects. 4606 // Therefore, tell all non-raw users to re-optimize themselves, 4607 // after skipping the memory effects of this initialization. 4608 PhaseIterGVN* igvn = phase->is_IterGVN(); 4609 if (igvn) igvn->add_users_to_worklist(this); 4610 } 4611 4612 // convenience function 4613 // return false if the init contains any stores already 4614 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 4615 InitializeNode* init = initialization(); 4616 if (init == nullptr || init->is_complete()) return false; 4617 init->remove_extra_zeroes(); 4618 // for now, if this allocation has already collected any inits, bail: 4619 if (init->is_non_zero()) return false; 4620 init->set_complete(phase); 4621 return true; 4622 } 4623 4624 void InitializeNode::remove_extra_zeroes() { 4625 if (req() == RawStores) return; 4626 Node* zmem = zero_memory(); 4627 uint fill = RawStores; 4628 for (uint i = fill; i < req(); i++) { 4629 Node* n = in(i); 4630 if (n->is_top() || n == zmem) continue; // skip 4631 if (fill < i) set_req(fill, n); // compact 4632 ++fill; 4633 } 4634 // delete any empty spaces created: 4635 while (fill < req()) { 4636 del_req(fill); 4637 } 4638 } 4639 4640 // Helper for remembering which stores go with which offsets. 4641 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) { 4642 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 4643 intptr_t offset = -1; 4644 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 4645 phase, offset); 4646 if (base == nullptr) return -1; // something is dead, 4647 if (offset < 0) return -1; // dead, dead 4648 return offset; 4649 } 4650 4651 // Helper for proving that an initialization expression is 4652 // "simple enough" to be folded into an object initialization. 4653 // Attempts to prove that a store's initial value 'n' can be captured 4654 // within the initialization without creating a vicious cycle, such as: 4655 // { Foo p = new Foo(); p.next = p; } 4656 // True for constants and parameters and small combinations thereof. 4657 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) { 4658 ResourceMark rm; 4659 Unique_Node_List worklist; 4660 worklist.push(value); 4661 4662 uint complexity_limit = 20; 4663 for (uint j = 0; j < worklist.size(); j++) { 4664 if (j >= complexity_limit) { 4665 return false; // Bail out if processed too many nodes 4666 } 4667 4668 Node* n = worklist.at(j); 4669 if (n == nullptr) continue; // (can this really happen?) 4670 if (n->is_Proj()) n = n->in(0); 4671 if (n == this) return false; // found a cycle 4672 if (n->is_Con()) continue; 4673 if (n->is_Start()) continue; // params, etc., are OK 4674 if (n->is_Root()) continue; // even better 4675 4676 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation 4677 if (n->is_CFG() && phase->is_dominator(n, allocation())) { 4678 continue; 4679 } 4680 4681 Node* ctl = n->in(0); 4682 if (ctl != nullptr && !ctl->is_top()) { 4683 if (ctl->is_Proj()) ctl = ctl->in(0); 4684 if (ctl == this) return false; 4685 4686 // If we already know that the enclosing memory op is pinned right after 4687 // the init, then any control flow that the store has picked up 4688 // must have preceded the init, or else be equal to the init. 4689 // Even after loop optimizations (which might change control edges) 4690 // a store is never pinned *before* the availability of its inputs. 4691 if (!MemNode::all_controls_dominate(n, this)) { 4692 return false; // failed to prove a good control 4693 } 4694 } 4695 4696 // Check data edges for possible dependencies on 'this'. 4697 for (uint i = 1; i < n->req(); i++) { 4698 Node* m = n->in(i); 4699 if (m == nullptr || m == n || m->is_top()) continue; 4700 4701 // Only process data inputs once 4702 worklist.push(m); 4703 } 4704 } 4705 4706 return true; 4707 } 4708 4709 // Here are all the checks a Store must pass before it can be moved into 4710 // an initialization. Returns zero if a check fails. 4711 // On success, returns the (constant) offset to which the store applies, 4712 // within the initialized memory. 4713 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) { 4714 const int FAIL = 0; 4715 if (st->req() != MemNode::ValueIn + 1) 4716 return FAIL; // an inscrutable StoreNode (card mark?) 4717 Node* ctl = st->in(MemNode::Control); 4718 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this)) 4719 return FAIL; // must be unconditional after the initialization 4720 Node* mem = st->in(MemNode::Memory); 4721 if (!(mem->is_Proj() && mem->in(0) == this)) 4722 return FAIL; // must not be preceded by other stores 4723 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4724 if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) && 4725 !bs->can_initialize_object(st)) { 4726 return FAIL; 4727 } 4728 Node* adr = st->in(MemNode::Address); 4729 intptr_t offset; 4730 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 4731 if (alloc == nullptr) 4732 return FAIL; // inscrutable address 4733 if (alloc != allocation()) 4734 return FAIL; // wrong allocation! (store needs to float up) 4735 int size_in_bytes = st->memory_size(); 4736 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) { 4737 return FAIL; // mismatched access 4738 } 4739 Node* val = st->in(MemNode::ValueIn); 4740 4741 if (!detect_init_independence(val, phase)) 4742 return FAIL; // stored value must be 'simple enough' 4743 4744 // The Store can be captured only if nothing after the allocation 4745 // and before the Store is using the memory location that the store 4746 // overwrites. 4747 bool failed = false; 4748 // If is_complete_with_arraycopy() is true the shape of the graph is 4749 // well defined and is safe so no need for extra checks. 4750 if (!is_complete_with_arraycopy()) { 4751 // We are going to look at each use of the memory state following 4752 // the allocation to make sure nothing reads the memory that the 4753 // Store writes. 4754 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 4755 int alias_idx = phase->C->get_alias_index(t_adr); 4756 ResourceMark rm; 4757 Unique_Node_List mems; 4758 mems.push(mem); 4759 Node* unique_merge = nullptr; 4760 for (uint next = 0; next < mems.size(); ++next) { 4761 Node *m = mems.at(next); 4762 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 4763 Node *n = m->fast_out(j); 4764 if (n->outcnt() == 0) { 4765 continue; 4766 } 4767 if (n == st) { 4768 continue; 4769 } else if (n->in(0) != nullptr && n->in(0) != ctl) { 4770 // If the control of this use is different from the control 4771 // of the Store which is right after the InitializeNode then 4772 // this node cannot be between the InitializeNode and the 4773 // Store. 4774 continue; 4775 } else if (n->is_MergeMem()) { 4776 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 4777 // We can hit a MergeMemNode (that will likely go away 4778 // later) that is a direct use of the memory state 4779 // following the InitializeNode on the same slice as the 4780 // store node that we'd like to capture. We need to check 4781 // the uses of the MergeMemNode. 4782 mems.push(n); 4783 } 4784 } else if (n->is_Mem()) { 4785 Node* other_adr = n->in(MemNode::Address); 4786 if (other_adr == adr) { 4787 failed = true; 4788 break; 4789 } else { 4790 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 4791 if (other_t_adr != nullptr) { 4792 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 4793 if (other_alias_idx == alias_idx) { 4794 // A load from the same memory slice as the store right 4795 // after the InitializeNode. We check the control of the 4796 // object/array that is loaded from. If it's the same as 4797 // the store control then we cannot capture the store. 4798 assert(!n->is_Store(), "2 stores to same slice on same control?"); 4799 Node* base = other_adr; 4800 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 4801 base = base->in(AddPNode::Base); 4802 if (base != nullptr) { 4803 base = base->uncast(); 4804 if (base->is_Proj() && base->in(0) == alloc) { 4805 failed = true; 4806 break; 4807 } 4808 } 4809 } 4810 } 4811 } 4812 } else { 4813 failed = true; 4814 break; 4815 } 4816 } 4817 } 4818 } 4819 if (failed) { 4820 if (!can_reshape) { 4821 // We decided we couldn't capture the store during parsing. We 4822 // should try again during the next IGVN once the graph is 4823 // cleaner. 4824 phase->C->record_for_igvn(st); 4825 } 4826 return FAIL; 4827 } 4828 4829 return offset; // success 4830 } 4831 4832 // Find the captured store in(i) which corresponds to the range 4833 // [start..start+size) in the initialized object. 4834 // If there is one, return its index i. If there isn't, return the 4835 // negative of the index where it should be inserted. 4836 // Return 0 if the queried range overlaps an initialization boundary 4837 // or if dead code is encountered. 4838 // If size_in_bytes is zero, do not bother with overlap checks. 4839 int InitializeNode::captured_store_insertion_point(intptr_t start, 4840 int size_in_bytes, 4841 PhaseValues* phase) { 4842 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize); 4843 4844 if (is_complete()) 4845 return FAIL; // arraycopy got here first; punt 4846 4847 assert(allocation() != nullptr, "must be present"); 4848 4849 // no negatives, no header fields: 4850 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 4851 4852 // after a certain size, we bail out on tracking all the stores: 4853 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 4854 if (start >= ti_limit) return FAIL; 4855 4856 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 4857 if (i >= limit) return -(int)i; // not found; here is where to put it 4858 4859 Node* st = in(i); 4860 intptr_t st_off = get_store_offset(st, phase); 4861 if (st_off < 0) { 4862 if (st != zero_memory()) { 4863 return FAIL; // bail out if there is dead garbage 4864 } 4865 } else if (st_off > start) { 4866 // ...we are done, since stores are ordered 4867 if (st_off < start + size_in_bytes) { 4868 return FAIL; // the next store overlaps 4869 } 4870 return -(int)i; // not found; here is where to put it 4871 } else if (st_off < start) { 4872 assert(st->as_Store()->memory_size() <= MAX_STORE, ""); 4873 if (size_in_bytes != 0 && 4874 start < st_off + MAX_STORE && 4875 start < st_off + st->as_Store()->memory_size()) { 4876 return FAIL; // the previous store overlaps 4877 } 4878 } else { 4879 if (size_in_bytes != 0 && 4880 st->as_Store()->memory_size() != size_in_bytes) { 4881 return FAIL; // mismatched store size 4882 } 4883 return i; 4884 } 4885 4886 ++i; 4887 } 4888 } 4889 4890 // Look for a captured store which initializes at the offset 'start' 4891 // with the given size. If there is no such store, and no other 4892 // initialization interferes, then return zero_memory (the memory 4893 // projection of the AllocateNode). 4894 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 4895 PhaseValues* phase) { 4896 assert(stores_are_sane(phase), ""); 4897 int i = captured_store_insertion_point(start, size_in_bytes, phase); 4898 if (i == 0) { 4899 return nullptr; // something is dead 4900 } else if (i < 0) { 4901 return zero_memory(); // just primordial zero bits here 4902 } else { 4903 Node* st = in(i); // here is the store at this position 4904 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 4905 return st; 4906 } 4907 } 4908 4909 // Create, as a raw pointer, an address within my new object at 'offset'. 4910 Node* InitializeNode::make_raw_address(intptr_t offset, 4911 PhaseGVN* phase) { 4912 Node* addr = in(RawAddress); 4913 if (offset != 0) { 4914 Compile* C = phase->C; 4915 addr = phase->transform( new AddPNode(C->top(), addr, 4916 phase->MakeConX(offset)) ); 4917 } 4918 return addr; 4919 } 4920 4921 // Clone the given store, converting it into a raw store 4922 // initializing a field or element of my new object. 4923 // Caller is responsible for retiring the original store, 4924 // with subsume_node or the like. 4925 // 4926 // From the example above InitializeNode::InitializeNode, 4927 // here are the old stores to be captured: 4928 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 4929 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 4930 // 4931 // Here is the changed code; note the extra edges on init: 4932 // alloc = (Allocate ...) 4933 // rawoop = alloc.RawAddress 4934 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 4935 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 4936 // init = (Initialize alloc.Control alloc.Memory rawoop 4937 // rawstore1 rawstore2) 4938 // 4939 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 4940 PhaseGVN* phase, bool can_reshape) { 4941 assert(stores_are_sane(phase), ""); 4942 4943 if (start < 0) return nullptr; 4944 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 4945 4946 Compile* C = phase->C; 4947 int size_in_bytes = st->memory_size(); 4948 int i = captured_store_insertion_point(start, size_in_bytes, phase); 4949 if (i == 0) return nullptr; // bail out 4950 Node* prev_mem = nullptr; // raw memory for the captured store 4951 if (i > 0) { 4952 prev_mem = in(i); // there is a pre-existing store under this one 4953 set_req(i, C->top()); // temporarily disconnect it 4954 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 4955 } else { 4956 i = -i; // no pre-existing store 4957 prev_mem = zero_memory(); // a slice of the newly allocated object 4958 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 4959 set_req(--i, C->top()); // reuse this edge; it has been folded away 4960 else 4961 ins_req(i, C->top()); // build a new edge 4962 } 4963 Node* new_st = st->clone(); 4964 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4965 new_st->set_req(MemNode::Control, in(Control)); 4966 new_st->set_req(MemNode::Memory, prev_mem); 4967 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 4968 bs->eliminate_gc_barrier_data(new_st); 4969 new_st = phase->transform(new_st); 4970 4971 // At this point, new_st might have swallowed a pre-existing store 4972 // at the same offset, or perhaps new_st might have disappeared, 4973 // if it redundantly stored the same value (or zero to fresh memory). 4974 4975 // In any case, wire it in: 4976 PhaseIterGVN* igvn = phase->is_IterGVN(); 4977 if (igvn) { 4978 igvn->rehash_node_delayed(this); 4979 } 4980 set_req(i, new_st); 4981 4982 // The caller may now kill the old guy. 4983 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 4984 assert(check_st == new_st || check_st == nullptr, "must be findable"); 4985 assert(!is_complete(), ""); 4986 return new_st; 4987 } 4988 4989 static bool store_constant(jlong* tiles, int num_tiles, 4990 intptr_t st_off, int st_size, 4991 jlong con) { 4992 if ((st_off & (st_size-1)) != 0) 4993 return false; // strange store offset (assume size==2**N) 4994 address addr = (address)tiles + st_off; 4995 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 4996 switch (st_size) { 4997 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 4998 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 4999 case sizeof(jint): *(jint*) addr = (jint) con; break; 5000 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 5001 default: return false; // strange store size (detect size!=2**N here) 5002 } 5003 return true; // return success to caller 5004 } 5005 5006 // Coalesce subword constants into int constants and possibly 5007 // into long constants. The goal, if the CPU permits, 5008 // is to initialize the object with a small number of 64-bit tiles. 5009 // Also, convert floating-point constants to bit patterns. 5010 // Non-constants are not relevant to this pass. 5011 // 5012 // In terms of the running example on InitializeNode::InitializeNode 5013 // and InitializeNode::capture_store, here is the transformation 5014 // of rawstore1 and rawstore2 into rawstore12: 5015 // alloc = (Allocate ...) 5016 // rawoop = alloc.RawAddress 5017 // tile12 = 0x00010002 5018 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 5019 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 5020 // 5021 void 5022 InitializeNode::coalesce_subword_stores(intptr_t header_size, 5023 Node* size_in_bytes, 5024 PhaseGVN* phase) { 5025 Compile* C = phase->C; 5026 5027 assert(stores_are_sane(phase), ""); 5028 // Note: After this pass, they are not completely sane, 5029 // since there may be some overlaps. 5030 5031 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 5032 5033 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 5034 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 5035 size_limit = MIN2(size_limit, ti_limit); 5036 size_limit = align_up(size_limit, BytesPerLong); 5037 int num_tiles = size_limit / BytesPerLong; 5038 5039 // allocate space for the tile map: 5040 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 5041 jlong tiles_buf[small_len]; 5042 Node* nodes_buf[small_len]; 5043 jlong inits_buf[small_len]; 5044 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 5045 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 5046 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 5047 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 5048 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 5049 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 5050 // tiles: exact bitwise model of all primitive constants 5051 // nodes: last constant-storing node subsumed into the tiles model 5052 // inits: which bytes (in each tile) are touched by any initializations 5053 5054 //// Pass A: Fill in the tile model with any relevant stores. 5055 5056 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 5057 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 5058 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 5059 Node* zmem = zero_memory(); // initially zero memory state 5060 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 5061 Node* st = in(i); 5062 intptr_t st_off = get_store_offset(st, phase); 5063 5064 // Figure out the store's offset and constant value: 5065 if (st_off < header_size) continue; //skip (ignore header) 5066 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 5067 int st_size = st->as_Store()->memory_size(); 5068 if (st_off + st_size > size_limit) break; 5069 5070 // Record which bytes are touched, whether by constant or not. 5071 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 5072 continue; // skip (strange store size) 5073 5074 const Type* val = phase->type(st->in(MemNode::ValueIn)); 5075 if (!val->singleton()) continue; //skip (non-con store) 5076 BasicType type = val->basic_type(); 5077 5078 jlong con = 0; 5079 switch (type) { 5080 case T_INT: con = val->is_int()->get_con(); break; 5081 case T_LONG: con = val->is_long()->get_con(); break; 5082 case T_FLOAT: con = jint_cast(val->getf()); break; 5083 case T_DOUBLE: con = jlong_cast(val->getd()); break; 5084 default: continue; //skip (odd store type) 5085 } 5086 5087 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 5088 st->Opcode() == Op_StoreL) { 5089 continue; // This StoreL is already optimal. 5090 } 5091 5092 // Store down the constant. 5093 store_constant(tiles, num_tiles, st_off, st_size, con); 5094 5095 intptr_t j = st_off >> LogBytesPerLong; 5096 5097 if (type == T_INT && st_size == BytesPerInt 5098 && (st_off & BytesPerInt) == BytesPerInt) { 5099 jlong lcon = tiles[j]; 5100 if (!Matcher::isSimpleConstant64(lcon) && 5101 st->Opcode() == Op_StoreI) { 5102 // This StoreI is already optimal by itself. 5103 jint* intcon = (jint*) &tiles[j]; 5104 intcon[1] = 0; // undo the store_constant() 5105 5106 // If the previous store is also optimal by itself, back up and 5107 // undo the action of the previous loop iteration... if we can. 5108 // But if we can't, just let the previous half take care of itself. 5109 st = nodes[j]; 5110 st_off -= BytesPerInt; 5111 con = intcon[0]; 5112 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) { 5113 assert(st_off >= header_size, "still ignoring header"); 5114 assert(get_store_offset(st, phase) == st_off, "must be"); 5115 assert(in(i-1) == zmem, "must be"); 5116 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 5117 assert(con == tcon->is_int()->get_con(), "must be"); 5118 // Undo the effects of the previous loop trip, which swallowed st: 5119 intcon[0] = 0; // undo store_constant() 5120 set_req(i-1, st); // undo set_req(i, zmem) 5121 nodes[j] = nullptr; // undo nodes[j] = st 5122 --old_subword; // undo ++old_subword 5123 } 5124 continue; // This StoreI is already optimal. 5125 } 5126 } 5127 5128 // This store is not needed. 5129 set_req(i, zmem); 5130 nodes[j] = st; // record for the moment 5131 if (st_size < BytesPerLong) // something has changed 5132 ++old_subword; // includes int/float, but who's counting... 5133 else ++old_long; 5134 } 5135 5136 if ((old_subword + old_long) == 0) 5137 return; // nothing more to do 5138 5139 //// Pass B: Convert any non-zero tiles into optimal constant stores. 5140 // Be sure to insert them before overlapping non-constant stores. 5141 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 5142 for (int j = 0; j < num_tiles; j++) { 5143 jlong con = tiles[j]; 5144 jlong init = inits[j]; 5145 if (con == 0) continue; 5146 jint con0, con1; // split the constant, address-wise 5147 jint init0, init1; // split the init map, address-wise 5148 { union { jlong con; jint intcon[2]; } u; 5149 u.con = con; 5150 con0 = u.intcon[0]; 5151 con1 = u.intcon[1]; 5152 u.con = init; 5153 init0 = u.intcon[0]; 5154 init1 = u.intcon[1]; 5155 } 5156 5157 Node* old = nodes[j]; 5158 assert(old != nullptr, "need the prior store"); 5159 intptr_t offset = (j * BytesPerLong); 5160 5161 bool split = !Matcher::isSimpleConstant64(con); 5162 5163 if (offset < header_size) { 5164 assert(offset + BytesPerInt >= header_size, "second int counts"); 5165 assert(*(jint*)&tiles[j] == 0, "junk in header"); 5166 split = true; // only the second word counts 5167 // Example: int a[] = { 42 ... } 5168 } else if (con0 == 0 && init0 == -1) { 5169 split = true; // first word is covered by full inits 5170 // Example: int a[] = { ... foo(), 42 ... } 5171 } else if (con1 == 0 && init1 == -1) { 5172 split = true; // second word is covered by full inits 5173 // Example: int a[] = { ... 42, foo() ... } 5174 } 5175 5176 // Here's a case where init0 is neither 0 nor -1: 5177 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 5178 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 5179 // In this case the tile is not split; it is (jlong)42. 5180 // The big tile is stored down, and then the foo() value is inserted. 5181 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 5182 5183 Node* ctl = old->in(MemNode::Control); 5184 Node* adr = make_raw_address(offset, phase); 5185 const TypePtr* atp = TypeRawPtr::BOTTOM; 5186 5187 // One or two coalesced stores to plop down. 5188 Node* st[2]; 5189 intptr_t off[2]; 5190 int nst = 0; 5191 if (!split) { 5192 ++new_long; 5193 off[nst] = offset; 5194 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 5195 phase->longcon(con), T_LONG, MemNode::unordered); 5196 } else { 5197 // Omit either if it is a zero. 5198 if (con0 != 0) { 5199 ++new_int; 5200 off[nst] = offset; 5201 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 5202 phase->intcon(con0), T_INT, MemNode::unordered); 5203 } 5204 if (con1 != 0) { 5205 ++new_int; 5206 offset += BytesPerInt; 5207 adr = make_raw_address(offset, phase); 5208 off[nst] = offset; 5209 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 5210 phase->intcon(con1), T_INT, MemNode::unordered); 5211 } 5212 } 5213 5214 // Insert second store first, then the first before the second. 5215 // Insert each one just before any overlapping non-constant stores. 5216 while (nst > 0) { 5217 Node* st1 = st[--nst]; 5218 C->copy_node_notes_to(st1, old); 5219 st1 = phase->transform(st1); 5220 offset = off[nst]; 5221 assert(offset >= header_size, "do not smash header"); 5222 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 5223 guarantee(ins_idx != 0, "must re-insert constant store"); 5224 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 5225 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 5226 set_req(--ins_idx, st1); 5227 else 5228 ins_req(ins_idx, st1); 5229 } 5230 } 5231 5232 if (PrintCompilation && WizardMode) 5233 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 5234 old_subword, old_long, new_int, new_long); 5235 if (C->log() != nullptr) 5236 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 5237 old_subword, old_long, new_int, new_long); 5238 5239 // Clean up any remaining occurrences of zmem: 5240 remove_extra_zeroes(); 5241 } 5242 5243 // Explore forward from in(start) to find the first fully initialized 5244 // word, and return its offset. Skip groups of subword stores which 5245 // together initialize full words. If in(start) is itself part of a 5246 // fully initialized word, return the offset of in(start). If there 5247 // are no following full-word stores, or if something is fishy, return 5248 // a negative value. 5249 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 5250 int int_map = 0; 5251 intptr_t int_map_off = 0; 5252 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 5253 5254 for (uint i = start, limit = req(); i < limit; i++) { 5255 Node* st = in(i); 5256 5257 intptr_t st_off = get_store_offset(st, phase); 5258 if (st_off < 0) break; // return conservative answer 5259 5260 int st_size = st->as_Store()->memory_size(); 5261 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 5262 return st_off; // we found a complete word init 5263 } 5264 5265 // update the map: 5266 5267 intptr_t this_int_off = align_down(st_off, BytesPerInt); 5268 if (this_int_off != int_map_off) { 5269 // reset the map: 5270 int_map = 0; 5271 int_map_off = this_int_off; 5272 } 5273 5274 int subword_off = st_off - this_int_off; 5275 int_map |= right_n_bits(st_size) << subword_off; 5276 if ((int_map & FULL_MAP) == FULL_MAP) { 5277 return this_int_off; // we found a complete word init 5278 } 5279 5280 // Did this store hit or cross the word boundary? 5281 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt); 5282 if (next_int_off == this_int_off + BytesPerInt) { 5283 // We passed the current int, without fully initializing it. 5284 int_map_off = next_int_off; 5285 int_map >>= BytesPerInt; 5286 } else if (next_int_off > this_int_off + BytesPerInt) { 5287 // We passed the current and next int. 5288 return this_int_off + BytesPerInt; 5289 } 5290 } 5291 5292 return -1; 5293 } 5294 5295 5296 // Called when the associated AllocateNode is expanded into CFG. 5297 // At this point, we may perform additional optimizations. 5298 // Linearize the stores by ascending offset, to make memory 5299 // activity as coherent as possible. 5300 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 5301 intptr_t header_size, 5302 Node* size_in_bytes, 5303 PhaseIterGVN* phase) { 5304 assert(!is_complete(), "not already complete"); 5305 assert(stores_are_sane(phase), ""); 5306 assert(allocation() != nullptr, "must be present"); 5307 5308 remove_extra_zeroes(); 5309 5310 if (ReduceFieldZeroing || ReduceBulkZeroing) 5311 // reduce instruction count for common initialization patterns 5312 coalesce_subword_stores(header_size, size_in_bytes, phase); 5313 5314 Node* zmem = zero_memory(); // initially zero memory state 5315 Node* inits = zmem; // accumulating a linearized chain of inits 5316 #ifdef ASSERT 5317 intptr_t first_offset = allocation()->minimum_header_size(); 5318 intptr_t last_init_off = first_offset; // previous init offset 5319 intptr_t last_init_end = first_offset; // previous init offset+size 5320 intptr_t last_tile_end = first_offset; // previous tile offset+size 5321 #endif 5322 intptr_t zeroes_done = header_size; 5323 5324 bool do_zeroing = true; // we might give up if inits are very sparse 5325 int big_init_gaps = 0; // how many large gaps have we seen? 5326 5327 if (UseTLAB && ZeroTLAB) do_zeroing = false; 5328 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 5329 5330 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 5331 Node* st = in(i); 5332 intptr_t st_off = get_store_offset(st, phase); 5333 if (st_off < 0) 5334 break; // unknown junk in the inits 5335 if (st->in(MemNode::Memory) != zmem) 5336 break; // complicated store chains somehow in list 5337 5338 int st_size = st->as_Store()->memory_size(); 5339 intptr_t next_init_off = st_off + st_size; 5340 5341 if (do_zeroing && zeroes_done < next_init_off) { 5342 // See if this store needs a zero before it or under it. 5343 intptr_t zeroes_needed = st_off; 5344 5345 if (st_size < BytesPerInt) { 5346 // Look for subword stores which only partially initialize words. 5347 // If we find some, we must lay down some word-level zeroes first, 5348 // underneath the subword stores. 5349 // 5350 // Examples: 5351 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 5352 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 5353 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 5354 // 5355 // Note: coalesce_subword_stores may have already done this, 5356 // if it was prompted by constant non-zero subword initializers. 5357 // But this case can still arise with non-constant stores. 5358 5359 intptr_t next_full_store = find_next_fullword_store(i, phase); 5360 5361 // In the examples above: 5362 // in(i) p q r s x y z 5363 // st_off 12 13 14 15 12 13 14 5364 // st_size 1 1 1 1 1 1 1 5365 // next_full_s. 12 16 16 16 16 16 16 5366 // z's_done 12 16 16 16 12 16 12 5367 // z's_needed 12 16 16 16 16 16 16 5368 // zsize 0 0 0 0 4 0 4 5369 if (next_full_store < 0) { 5370 // Conservative tack: Zero to end of current word. 5371 zeroes_needed = align_up(zeroes_needed, BytesPerInt); 5372 } else { 5373 // Zero to beginning of next fully initialized word. 5374 // Or, don't zero at all, if we are already in that word. 5375 assert(next_full_store >= zeroes_needed, "must go forward"); 5376 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 5377 zeroes_needed = next_full_store; 5378 } 5379 } 5380 5381 if (zeroes_needed > zeroes_done) { 5382 intptr_t zsize = zeroes_needed - zeroes_done; 5383 // Do some incremental zeroing on rawmem, in parallel with inits. 5384 zeroes_done = align_down(zeroes_done, BytesPerInt); 5385 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 5386 zeroes_done, zeroes_needed, 5387 phase); 5388 zeroes_done = zeroes_needed; 5389 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 5390 do_zeroing = false; // leave the hole, next time 5391 } 5392 } 5393 5394 // Collect the store and move on: 5395 phase->replace_input_of(st, MemNode::Memory, inits); 5396 inits = st; // put it on the linearized chain 5397 set_req(i, zmem); // unhook from previous position 5398 5399 if (zeroes_done == st_off) 5400 zeroes_done = next_init_off; 5401 5402 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 5403 5404 #ifdef ASSERT 5405 // Various order invariants. Weaker than stores_are_sane because 5406 // a large constant tile can be filled in by smaller non-constant stores. 5407 assert(st_off >= last_init_off, "inits do not reverse"); 5408 last_init_off = st_off; 5409 const Type* val = nullptr; 5410 if (st_size >= BytesPerInt && 5411 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 5412 (int)val->basic_type() < (int)T_OBJECT) { 5413 assert(st_off >= last_tile_end, "tiles do not overlap"); 5414 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 5415 last_tile_end = MAX2(last_tile_end, next_init_off); 5416 } else { 5417 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong); 5418 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 5419 assert(st_off >= last_init_end, "inits do not overlap"); 5420 last_init_end = next_init_off; // it's a non-tile 5421 } 5422 #endif //ASSERT 5423 } 5424 5425 remove_extra_zeroes(); // clear out all the zmems left over 5426 add_req(inits); 5427 5428 if (!(UseTLAB && ZeroTLAB)) { 5429 // If anything remains to be zeroed, zero it all now. 5430 zeroes_done = align_down(zeroes_done, BytesPerInt); 5431 // if it is the last unused 4 bytes of an instance, forget about it 5432 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 5433 if (zeroes_done + BytesPerLong >= size_limit) { 5434 AllocateNode* alloc = allocation(); 5435 assert(alloc != nullptr, "must be present"); 5436 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) { 5437 Node* klass_node = alloc->in(AllocateNode::KlassNode); 5438 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass(); 5439 if (zeroes_done == k->layout_helper()) 5440 zeroes_done = size_limit; 5441 } 5442 } 5443 if (zeroes_done < size_limit) { 5444 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 5445 zeroes_done, size_in_bytes, phase); 5446 } 5447 } 5448 5449 set_complete(phase); 5450 return rawmem; 5451 } 5452 5453 5454 #ifdef ASSERT 5455 bool InitializeNode::stores_are_sane(PhaseValues* phase) { 5456 if (is_complete()) 5457 return true; // stores could be anything at this point 5458 assert(allocation() != nullptr, "must be present"); 5459 intptr_t last_off = allocation()->minimum_header_size(); 5460 for (uint i = InitializeNode::RawStores; i < req(); i++) { 5461 Node* st = in(i); 5462 intptr_t st_off = get_store_offset(st, phase); 5463 if (st_off < 0) continue; // ignore dead garbage 5464 if (last_off > st_off) { 5465 tty->print_cr("*** bad store offset at %d: %zd > %zd", i, last_off, st_off); 5466 this->dump(2); 5467 assert(false, "ascending store offsets"); 5468 return false; 5469 } 5470 last_off = st_off + st->as_Store()->memory_size(); 5471 } 5472 return true; 5473 } 5474 #endif //ASSERT 5475 5476 5477 5478 5479 //============================MergeMemNode===================================== 5480 // 5481 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 5482 // contributing store or call operations. Each contributor provides the memory 5483 // state for a particular "alias type" (see Compile::alias_type). For example, 5484 // if a MergeMem has an input X for alias category #6, then any memory reference 5485 // to alias category #6 may use X as its memory state input, as an exact equivalent 5486 // to using the MergeMem as a whole. 5487 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 5488 // 5489 // (Here, the <N> notation gives the index of the relevant adr_type.) 5490 // 5491 // In one special case (and more cases in the future), alias categories overlap. 5492 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 5493 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 5494 // it is exactly equivalent to that state W: 5495 // MergeMem(<Bot>: W) <==> W 5496 // 5497 // Usually, the merge has more than one input. In that case, where inputs 5498 // overlap (i.e., one is Bot), the narrower alias type determines the memory 5499 // state for that type, and the wider alias type (Bot) fills in everywhere else: 5500 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 5501 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 5502 // 5503 // A merge can take a "wide" memory state as one of its narrow inputs. 5504 // This simply means that the merge observes out only the relevant parts of 5505 // the wide input. That is, wide memory states arriving at narrow merge inputs 5506 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 5507 // 5508 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 5509 // and that memory slices "leak through": 5510 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 5511 // 5512 // But, in such a cascade, repeated memory slices can "block the leak": 5513 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 5514 // 5515 // In the last example, Y is not part of the combined memory state of the 5516 // outermost MergeMem. The system must, of course, prevent unschedulable 5517 // memory states from arising, so you can be sure that the state Y is somehow 5518 // a precursor to state Y'. 5519 // 5520 // 5521 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 5522 // of each MergeMemNode array are exactly the numerical alias indexes, including 5523 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 5524 // Compile::alias_type (and kin) produce and manage these indexes. 5525 // 5526 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 5527 // (Note that this provides quick access to the top node inside MergeMem methods, 5528 // without the need to reach out via TLS to Compile::current.) 5529 // 5530 // As a consequence of what was just described, a MergeMem that represents a full 5531 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 5532 // containing all alias categories. 5533 // 5534 // MergeMem nodes never (?) have control inputs, so in(0) is null. 5535 // 5536 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 5537 // a memory state for the alias type <N>, or else the top node, meaning that 5538 // there is no particular input for that alias type. Note that the length of 5539 // a MergeMem is variable, and may be extended at any time to accommodate new 5540 // memory states at larger alias indexes. When merges grow, they are of course 5541 // filled with "top" in the unused in() positions. 5542 // 5543 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 5544 // (Top was chosen because it works smoothly with passes like GCM.) 5545 // 5546 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 5547 // the type of random VM bits like TLS references.) Since it is always the 5548 // first non-Bot memory slice, some low-level loops use it to initialize an 5549 // index variable: for (i = AliasIdxRaw; i < req(); i++). 5550 // 5551 // 5552 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 5553 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 5554 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 5555 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 5556 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 5557 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 5558 // 5559 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 5560 // really that different from the other memory inputs. An abbreviation called 5561 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 5562 // 5563 // 5564 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 5565 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 5566 // that "emerges though" the base memory will be marked as excluding the alias types 5567 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 5568 // 5569 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 5570 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 5571 // 5572 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 5573 // (It is currently unimplemented.) As you can see, the resulting merge is 5574 // actually a disjoint union of memory states, rather than an overlay. 5575 // 5576 5577 //------------------------------MergeMemNode----------------------------------- 5578 Node* MergeMemNode::make_empty_memory() { 5579 Node* empty_memory = (Node*) Compile::current()->top(); 5580 assert(empty_memory->is_top(), "correct sentinel identity"); 5581 return empty_memory; 5582 } 5583 5584 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 5585 init_class_id(Class_MergeMem); 5586 // all inputs are nullified in Node::Node(int) 5587 // set_input(0, nullptr); // no control input 5588 5589 // Initialize the edges uniformly to top, for starters. 5590 Node* empty_mem = make_empty_memory(); 5591 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 5592 init_req(i,empty_mem); 5593 } 5594 assert(empty_memory() == empty_mem, ""); 5595 5596 if( new_base != nullptr && new_base->is_MergeMem() ) { 5597 MergeMemNode* mdef = new_base->as_MergeMem(); 5598 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 5599 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 5600 mms.set_memory(mms.memory2()); 5601 } 5602 assert(base_memory() == mdef->base_memory(), ""); 5603 } else { 5604 set_base_memory(new_base); 5605 } 5606 } 5607 5608 // Make a new, untransformed MergeMem with the same base as 'mem'. 5609 // If mem is itself a MergeMem, populate the result with the same edges. 5610 MergeMemNode* MergeMemNode::make(Node* mem) { 5611 return new MergeMemNode(mem); 5612 } 5613 5614 //------------------------------cmp-------------------------------------------- 5615 uint MergeMemNode::hash() const { return NO_HASH; } 5616 bool MergeMemNode::cmp( const Node &n ) const { 5617 return (&n == this); // Always fail except on self 5618 } 5619 5620 //------------------------------Identity--------------------------------------- 5621 Node* MergeMemNode::Identity(PhaseGVN* phase) { 5622 // Identity if this merge point does not record any interesting memory 5623 // disambiguations. 5624 Node* base_mem = base_memory(); 5625 Node* empty_mem = empty_memory(); 5626 if (base_mem != empty_mem) { // Memory path is not dead? 5627 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 5628 Node* mem = in(i); 5629 if (mem != empty_mem && mem != base_mem) { 5630 return this; // Many memory splits; no change 5631 } 5632 } 5633 } 5634 return base_mem; // No memory splits; ID on the one true input 5635 } 5636 5637 //------------------------------Ideal------------------------------------------ 5638 // This method is invoked recursively on chains of MergeMem nodes 5639 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 5640 // Remove chain'd MergeMems 5641 // 5642 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 5643 // relative to the "in(Bot)". Since we are patching both at the same time, 5644 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 5645 // but rewrite each "in(i)" relative to the new "in(Bot)". 5646 Node *progress = nullptr; 5647 5648 5649 Node* old_base = base_memory(); 5650 Node* empty_mem = empty_memory(); 5651 if (old_base == empty_mem) 5652 return nullptr; // Dead memory path. 5653 5654 MergeMemNode* old_mbase; 5655 if (old_base != nullptr && old_base->is_MergeMem()) 5656 old_mbase = old_base->as_MergeMem(); 5657 else 5658 old_mbase = nullptr; 5659 Node* new_base = old_base; 5660 5661 // simplify stacked MergeMems in base memory 5662 if (old_mbase) new_base = old_mbase->base_memory(); 5663 5664 // the base memory might contribute new slices beyond my req() 5665 if (old_mbase) grow_to_match(old_mbase); 5666 5667 // Note: We do not call verify_sparse on entry, because inputs 5668 // can normalize to the base_memory via subsume_node or similar 5669 // mechanisms. This method repairs that damage. 5670 5671 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 5672 5673 // Look at each slice. 5674 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 5675 Node* old_in = in(i); 5676 // calculate the old memory value 5677 Node* old_mem = old_in; 5678 if (old_mem == empty_mem) old_mem = old_base; 5679 assert(old_mem == memory_at(i), ""); 5680 5681 // maybe update (reslice) the old memory value 5682 5683 // simplify stacked MergeMems 5684 Node* new_mem = old_mem; 5685 MergeMemNode* old_mmem; 5686 if (old_mem != nullptr && old_mem->is_MergeMem()) 5687 old_mmem = old_mem->as_MergeMem(); 5688 else 5689 old_mmem = nullptr; 5690 if (old_mmem == this) { 5691 // This can happen if loops break up and safepoints disappear. 5692 // A merge of BotPtr (default) with a RawPtr memory derived from a 5693 // safepoint can be rewritten to a merge of the same BotPtr with 5694 // the BotPtr phi coming into the loop. If that phi disappears 5695 // also, we can end up with a self-loop of the mergemem. 5696 // In general, if loops degenerate and memory effects disappear, 5697 // a mergemem can be left looking at itself. This simply means 5698 // that the mergemem's default should be used, since there is 5699 // no longer any apparent effect on this slice. 5700 // Note: If a memory slice is a MergeMem cycle, it is unreachable 5701 // from start. Update the input to TOP. 5702 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 5703 } 5704 else if (old_mmem != nullptr) { 5705 new_mem = old_mmem->memory_at(i); 5706 } 5707 // else preceding memory was not a MergeMem 5708 5709 // maybe store down a new value 5710 Node* new_in = new_mem; 5711 if (new_in == new_base) new_in = empty_mem; 5712 5713 if (new_in != old_in) { 5714 // Warning: Do not combine this "if" with the previous "if" 5715 // A memory slice might have be be rewritten even if it is semantically 5716 // unchanged, if the base_memory value has changed. 5717 set_req_X(i, new_in, phase); 5718 progress = this; // Report progress 5719 } 5720 } 5721 5722 if (new_base != old_base) { 5723 set_req_X(Compile::AliasIdxBot, new_base, phase); 5724 // Don't use set_base_memory(new_base), because we need to update du. 5725 assert(base_memory() == new_base, ""); 5726 progress = this; 5727 } 5728 5729 if( base_memory() == this ) { 5730 // a self cycle indicates this memory path is dead 5731 set_req(Compile::AliasIdxBot, empty_mem); 5732 } 5733 5734 // Resolve external cycles by calling Ideal on a MergeMem base_memory 5735 // Recursion must occur after the self cycle check above 5736 if( base_memory()->is_MergeMem() ) { 5737 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 5738 Node *m = phase->transform(new_mbase); // Rollup any cycles 5739 if( m != nullptr && 5740 (m->is_top() || 5741 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) { 5742 // propagate rollup of dead cycle to self 5743 set_req(Compile::AliasIdxBot, empty_mem); 5744 } 5745 } 5746 5747 if( base_memory() == empty_mem ) { 5748 progress = this; 5749 // Cut inputs during Parse phase only. 5750 // During Optimize phase a dead MergeMem node will be subsumed by Top. 5751 if( !can_reshape ) { 5752 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 5753 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 5754 } 5755 } 5756 } 5757 5758 if( !progress && base_memory()->is_Phi() && can_reshape ) { 5759 // Check if PhiNode::Ideal's "Split phis through memory merges" 5760 // transform should be attempted. Look for this->phi->this cycle. 5761 uint merge_width = req(); 5762 if (merge_width > Compile::AliasIdxRaw) { 5763 PhiNode* phi = base_memory()->as_Phi(); 5764 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 5765 if (phi->in(i) == this) { 5766 phase->is_IterGVN()->_worklist.push(phi); 5767 break; 5768 } 5769 } 5770 } 5771 } 5772 5773 assert(progress || verify_sparse(), "please, no dups of base"); 5774 return progress; 5775 } 5776 5777 //-------------------------set_base_memory------------------------------------- 5778 void MergeMemNode::set_base_memory(Node *new_base) { 5779 Node* empty_mem = empty_memory(); 5780 set_req(Compile::AliasIdxBot, new_base); 5781 assert(memory_at(req()) == new_base, "must set default memory"); 5782 // Clear out other occurrences of new_base: 5783 if (new_base != empty_mem) { 5784 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 5785 if (in(i) == new_base) set_req(i, empty_mem); 5786 } 5787 } 5788 } 5789 5790 //------------------------------out_RegMask------------------------------------ 5791 const RegMask &MergeMemNode::out_RegMask() const { 5792 return RegMask::Empty; 5793 } 5794 5795 //------------------------------dump_spec-------------------------------------- 5796 #ifndef PRODUCT 5797 void MergeMemNode::dump_spec(outputStream *st) const { 5798 st->print(" {"); 5799 Node* base_mem = base_memory(); 5800 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 5801 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem; 5802 if (mem == base_mem) { st->print(" -"); continue; } 5803 st->print( " N%d:", mem->_idx ); 5804 Compile::current()->get_adr_type(i)->dump_on(st); 5805 } 5806 st->print(" }"); 5807 } 5808 #endif // !PRODUCT 5809 5810 5811 #ifdef ASSERT 5812 static bool might_be_same(Node* a, Node* b) { 5813 if (a == b) return true; 5814 if (!(a->is_Phi() || b->is_Phi())) return false; 5815 // phis shift around during optimization 5816 return true; // pretty stupid... 5817 } 5818 5819 // verify a narrow slice (either incoming or outgoing) 5820 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 5821 if (!VerifyAliases) return; // don't bother to verify unless requested 5822 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error 5823 if (Node::in_dump()) return; // muzzle asserts when printing 5824 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 5825 assert(n != nullptr, ""); 5826 // Elide intervening MergeMem's 5827 while (n->is_MergeMem()) { 5828 n = n->as_MergeMem()->memory_at(alias_idx); 5829 } 5830 Compile* C = Compile::current(); 5831 const TypePtr* n_adr_type = n->adr_type(); 5832 if (n == m->empty_memory()) { 5833 // Implicit copy of base_memory() 5834 } else if (n_adr_type != TypePtr::BOTTOM) { 5835 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type"); 5836 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 5837 } else { 5838 // A few places like make_runtime_call "know" that VM calls are narrow, 5839 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 5840 bool expected_wide_mem = false; 5841 if (n == m->base_memory()) { 5842 expected_wide_mem = true; 5843 } else if (alias_idx == Compile::AliasIdxRaw || 5844 n == m->memory_at(Compile::AliasIdxRaw)) { 5845 expected_wide_mem = true; 5846 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 5847 // memory can "leak through" calls on channels that 5848 // are write-once. Allow this also. 5849 expected_wide_mem = true; 5850 } 5851 assert(expected_wide_mem, "expected narrow slice replacement"); 5852 } 5853 } 5854 #else // !ASSERT 5855 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 5856 #endif 5857 5858 5859 //-----------------------------memory_at--------------------------------------- 5860 Node* MergeMemNode::memory_at(uint alias_idx) const { 5861 assert(alias_idx >= Compile::AliasIdxRaw || 5862 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()), 5863 "must avoid base_memory and AliasIdxTop"); 5864 5865 // Otherwise, it is a narrow slice. 5866 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 5867 if (is_empty_memory(n)) { 5868 // the array is sparse; empty slots are the "top" node 5869 n = base_memory(); 5870 assert(Node::in_dump() 5871 || n == nullptr || n->bottom_type() == Type::TOP 5872 || n->adr_type() == nullptr // address is TOP 5873 || n->adr_type() == TypePtr::BOTTOM 5874 || n->adr_type() == TypeRawPtr::BOTTOM 5875 || !Compile::current()->do_aliasing(), 5876 "must be a wide memory"); 5877 // do_aliasing == false if we are organizing the memory states manually. 5878 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 5879 } else { 5880 // make sure the stored slice is sane 5881 #ifdef ASSERT 5882 if (VMError::is_error_reported() || Node::in_dump()) { 5883 } else if (might_be_same(n, base_memory())) { 5884 // Give it a pass: It is a mostly harmless repetition of the base. 5885 // This can arise normally from node subsumption during optimization. 5886 } else { 5887 verify_memory_slice(this, alias_idx, n); 5888 } 5889 #endif 5890 } 5891 return n; 5892 } 5893 5894 //---------------------------set_memory_at------------------------------------- 5895 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 5896 verify_memory_slice(this, alias_idx, n); 5897 Node* empty_mem = empty_memory(); 5898 if (n == base_memory()) n = empty_mem; // collapse default 5899 uint need_req = alias_idx+1; 5900 if (req() < need_req) { 5901 if (n == empty_mem) return; // already the default, so do not grow me 5902 // grow the sparse array 5903 do { 5904 add_req(empty_mem); 5905 } while (req() < need_req); 5906 } 5907 set_req( alias_idx, n ); 5908 } 5909 5910 5911 5912 //--------------------------iteration_setup------------------------------------ 5913 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 5914 if (other != nullptr) { 5915 grow_to_match(other); 5916 // invariant: the finite support of mm2 is within mm->req() 5917 #ifdef ASSERT 5918 for (uint i = req(); i < other->req(); i++) { 5919 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 5920 } 5921 #endif 5922 } 5923 // Replace spurious copies of base_memory by top. 5924 Node* base_mem = base_memory(); 5925 if (base_mem != nullptr && !base_mem->is_top()) { 5926 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 5927 if (in(i) == base_mem) 5928 set_req(i, empty_memory()); 5929 } 5930 } 5931 } 5932 5933 //---------------------------grow_to_match------------------------------------- 5934 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 5935 Node* empty_mem = empty_memory(); 5936 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 5937 // look for the finite support of the other memory 5938 for (uint i = other->req(); --i >= req(); ) { 5939 if (other->in(i) != empty_mem) { 5940 uint new_len = i+1; 5941 while (req() < new_len) add_req(empty_mem); 5942 break; 5943 } 5944 } 5945 } 5946 5947 //---------------------------verify_sparse------------------------------------- 5948 #ifndef PRODUCT 5949 bool MergeMemNode::verify_sparse() const { 5950 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 5951 Node* base_mem = base_memory(); 5952 // The following can happen in degenerate cases, since empty==top. 5953 if (is_empty_memory(base_mem)) return true; 5954 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 5955 assert(in(i) != nullptr, "sane slice"); 5956 if (in(i) == base_mem) return false; // should have been the sentinel value! 5957 } 5958 return true; 5959 } 5960 5961 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 5962 Node* n; 5963 n = mm->in(idx); 5964 if (mem == n) return true; // might be empty_memory() 5965 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 5966 if (mem == n) return true; 5967 return false; 5968 } 5969 #endif // !PRODUCT