1 /* 2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/ciFlatArrayKlass.hpp" 27 #include "classfile/javaClasses.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "compiler/compileLog.hpp" 30 #include "gc/shared/barrierSet.hpp" 31 #include "gc/shared/c2/barrierSetC2.hpp" 32 #include "gc/shared/tlab_globals.hpp" 33 #include "memory/allocation.inline.hpp" 34 #include "memory/resourceArea.hpp" 35 #include "oops/objArrayKlass.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/arraycopynode.hpp" 38 #include "opto/cfgnode.hpp" 39 #include "opto/regalloc.hpp" 40 #include "opto/compile.hpp" 41 #include "opto/connode.hpp" 42 #include "opto/convertnode.hpp" 43 #include "opto/inlinetypenode.hpp" 44 #include "opto/loopnode.hpp" 45 #include "opto/machnode.hpp" 46 #include "opto/matcher.hpp" 47 #include "opto/memnode.hpp" 48 #include "opto/mulnode.hpp" 49 #include "opto/narrowptrnode.hpp" 50 #include "opto/phaseX.hpp" 51 #include "opto/regmask.hpp" 52 #include "opto/rootnode.hpp" 53 #include "opto/vectornode.hpp" 54 #include "utilities/align.hpp" 55 #include "utilities/copy.hpp" 56 #include "utilities/macros.hpp" 57 #include "utilities/powerOfTwo.hpp" 58 #include "utilities/vmError.hpp" 59 60 // Portions of code courtesy of Clifford Click 61 62 // Optimization - Graph Style 63 64 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 65 66 //============================================================================= 67 uint MemNode::size_of() const { return sizeof(*this); } 68 69 const TypePtr *MemNode::adr_type() const { 70 Node* adr = in(Address); 71 if (adr == nullptr) return nullptr; // node is dead 72 const TypePtr* cross_check = nullptr; 73 DEBUG_ONLY(cross_check = _adr_type); 74 return calculate_adr_type(adr->bottom_type(), cross_check); 75 } 76 77 bool MemNode::check_if_adr_maybe_raw(Node* adr) { 78 if (adr != nullptr) { 79 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) { 80 return true; 81 } 82 } 83 return false; 84 } 85 86 #ifndef PRODUCT 87 void MemNode::dump_spec(outputStream *st) const { 88 if (in(Address) == nullptr) return; // node is dead 89 #ifndef ASSERT 90 // fake the missing field 91 const TypePtr* _adr_type = nullptr; 92 if (in(Address) != nullptr) 93 _adr_type = in(Address)->bottom_type()->isa_ptr(); 94 #endif 95 dump_adr_type(this, _adr_type, st); 96 97 Compile* C = Compile::current(); 98 if (C->alias_type(_adr_type)->is_volatile()) { 99 st->print(" Volatile!"); 100 } 101 if (_unaligned_access) { 102 st->print(" unaligned"); 103 } 104 if (_mismatched_access) { 105 st->print(" mismatched"); 106 } 107 if (_unsafe_access) { 108 st->print(" unsafe"); 109 } 110 } 111 112 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 113 st->print(" @"); 114 if (adr_type == nullptr) { 115 st->print("null"); 116 } else { 117 adr_type->dump_on(st); 118 Compile* C = Compile::current(); 119 Compile::AliasType* atp = nullptr; 120 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 121 if (atp == nullptr) 122 st->print(", idx=?\?;"); 123 else if (atp->index() == Compile::AliasIdxBot) 124 st->print(", idx=Bot;"); 125 else if (atp->index() == Compile::AliasIdxTop) 126 st->print(", idx=Top;"); 127 else if (atp->index() == Compile::AliasIdxRaw) 128 st->print(", idx=Raw;"); 129 else { 130 ciField* field = atp->field(); 131 if (field) { 132 st->print(", name="); 133 field->print_name_on(st); 134 } 135 st->print(", idx=%d;", atp->index()); 136 } 137 } 138 } 139 140 extern void print_alias_types(); 141 142 #endif 143 144 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 145 assert((t_oop != nullptr), "sanity"); 146 bool is_instance = t_oop->is_known_instance_field(); 147 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 148 (load != nullptr) && load->is_Load() && 149 (phase->is_IterGVN() != nullptr); 150 if (!(is_instance || is_boxed_value_load)) 151 return mchain; // don't try to optimize non-instance types 152 uint instance_id = t_oop->instance_id(); 153 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory); 154 Node *prev = nullptr; 155 Node *result = mchain; 156 while (prev != result) { 157 prev = result; 158 if (result == start_mem) 159 break; // hit one of our sentinels 160 // skip over a call which does not affect this memory slice 161 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 162 Node *proj_in = result->in(0); 163 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 164 break; // hit one of our sentinels 165 } else if (proj_in->is_Call()) { 166 // ArrayCopyNodes processed here as well 167 CallNode *call = proj_in->as_Call(); 168 if (!call->may_modify(t_oop, phase)) { // returns false for instances 169 result = call->in(TypeFunc::Memory); 170 } 171 } else if (proj_in->is_Initialize()) { 172 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 173 // Stop if this is the initialization for the object instance which 174 // contains this memory slice, otherwise skip over it. 175 if ((alloc == nullptr) || (alloc->_idx == instance_id)) { 176 break; 177 } 178 if (is_instance) { 179 result = proj_in->in(TypeFunc::Memory); 180 } else if (is_boxed_value_load) { 181 Node* klass = alloc->in(AllocateNode::KlassNode); 182 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 183 if (tklass->klass_is_exact() && !tklass->exact_klass()->equals(t_oop->is_instptr()->exact_klass())) { 184 result = proj_in->in(TypeFunc::Memory); // not related allocation 185 } 186 } 187 } else if (proj_in->is_MemBar()) { 188 ArrayCopyNode* ac = nullptr; 189 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) { 190 break; 191 } 192 result = proj_in->in(TypeFunc::Memory); 193 } else if (proj_in->is_top()) { 194 break; // dead code 195 } else { 196 assert(false, "unexpected projection"); 197 } 198 } else if (result->is_ClearArray()) { 199 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 200 // Can not bypass initialization of the instance 201 // we are looking for. 202 break; 203 } 204 // Otherwise skip it (the call updated 'result' value). 205 } else if (result->is_MergeMem()) { 206 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, nullptr, tty); 207 } 208 } 209 return result; 210 } 211 212 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 213 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 214 if (t_oop == nullptr) 215 return mchain; // don't try to optimize non-oop types 216 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 217 bool is_instance = t_oop->is_known_instance_field(); 218 PhaseIterGVN *igvn = phase->is_IterGVN(); 219 if (is_instance && igvn != nullptr && result->is_Phi()) { 220 PhiNode *mphi = result->as_Phi(); 221 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 222 const TypePtr *t = mphi->adr_type(); 223 bool do_split = false; 224 // In the following cases, Load memory input can be further optimized based on 225 // its precise address type 226 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) { 227 do_split = true; 228 } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) { 229 const TypeOopPtr* mem_t = 230 t->is_oopptr()->cast_to_exactness(true) 231 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 232 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()); 233 if (t_oop->is_aryptr()) { 234 mem_t = mem_t->is_aryptr() 235 ->cast_to_stable(t_oop->is_aryptr()->is_stable()) 236 ->cast_to_size(t_oop->is_aryptr()->size()) 237 ->cast_to_not_flat(t_oop->is_aryptr()->is_not_flat()) 238 ->cast_to_not_null_free(t_oop->is_aryptr()->is_not_null_free()) 239 ->with_offset(t_oop->is_aryptr()->offset()) 240 ->is_aryptr(); 241 } 242 do_split = mem_t == t_oop; 243 } 244 if (do_split) { 245 // clone the Phi with our address type 246 result = mphi->split_out_instance(t_adr, igvn); 247 } else { 248 // TODO 8325106 249 // assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 250 } 251 } 252 return result; 253 } 254 255 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 256 uint alias_idx = phase->C->get_alias_index(tp); 257 Node *mem = mmem; 258 #ifdef ASSERT 259 { 260 // Check that current type is consistent with the alias index used during graph construction 261 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 262 bool consistent = adr_check == nullptr || adr_check->empty() || 263 phase->C->must_alias(adr_check, alias_idx ); 264 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 265 if( !consistent && adr_check != nullptr && !adr_check->empty() && 266 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 267 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 268 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 269 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 270 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 271 // don't assert if it is dead code. 272 consistent = true; 273 } 274 if( !consistent ) { 275 st->print("alias_idx==%d, adr_check==", alias_idx); 276 if( adr_check == nullptr ) { 277 st->print("null"); 278 } else { 279 adr_check->dump(); 280 } 281 st->cr(); 282 print_alias_types(); 283 assert(consistent, "adr_check must match alias idx"); 284 } 285 } 286 #endif 287 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 288 // means an array I have not precisely typed yet. Do not do any 289 // alias stuff with it any time soon. 290 const TypeOopPtr *toop = tp->isa_oopptr(); 291 if (tp->base() != Type::AnyPtr && 292 !(toop && 293 toop->isa_instptr() && 294 toop->is_instptr()->instance_klass()->is_java_lang_Object() && 295 toop->offset() == Type::OffsetBot)) { 296 // compress paths and change unreachable cycles to TOP 297 // If not, we can update the input infinitely along a MergeMem cycle 298 // Equivalent code in PhiNode::Ideal 299 Node* m = phase->transform(mmem); 300 // If transformed to a MergeMem, get the desired slice 301 // Otherwise the returned node represents memory for every slice 302 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 303 // Update input if it is progress over what we have now 304 } 305 return mem; 306 } 307 308 //--------------------------Ideal_common--------------------------------------- 309 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 310 // Unhook non-raw memories from complete (macro-expanded) initializations. 311 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 312 // If our control input is a dead region, kill all below the region 313 Node *ctl = in(MemNode::Control); 314 if (ctl && remove_dead_region(phase, can_reshape)) 315 return this; 316 ctl = in(MemNode::Control); 317 // Don't bother trying to transform a dead node 318 if (ctl && ctl->is_top()) return NodeSentinel; 319 320 PhaseIterGVN *igvn = phase->is_IterGVN(); 321 // Wait if control on the worklist. 322 if (ctl && can_reshape && igvn != nullptr) { 323 Node* bol = nullptr; 324 Node* cmp = nullptr; 325 if (ctl->in(0)->is_If()) { 326 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 327 bol = ctl->in(0)->in(1); 328 if (bol->is_Bool()) 329 cmp = ctl->in(0)->in(1)->in(1); 330 } 331 if (igvn->_worklist.member(ctl) || 332 (bol != nullptr && igvn->_worklist.member(bol)) || 333 (cmp != nullptr && igvn->_worklist.member(cmp)) ) { 334 // This control path may be dead. 335 // Delay this memory node transformation until the control is processed. 336 igvn->_worklist.push(this); 337 return NodeSentinel; // caller will return null 338 } 339 } 340 // Ignore if memory is dead, or self-loop 341 Node *mem = in(MemNode::Memory); 342 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null 343 assert(mem != this, "dead loop in MemNode::Ideal"); 344 345 if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) { 346 // This memory slice may be dead. 347 // Delay this mem node transformation until the memory is processed. 348 igvn->_worklist.push(this); 349 return NodeSentinel; // caller will return null 350 } 351 352 Node *address = in(MemNode::Address); 353 const Type *t_adr = phase->type(address); 354 if (t_adr == Type::TOP) return NodeSentinel; // caller will return null 355 356 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) { 357 // Unsafe off-heap access with zero address. Remove access and other control users 358 // to not confuse optimizations and add a HaltNode to fail if this is ever executed. 359 assert(ctl != nullptr, "unsafe accesses should be control dependent"); 360 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) { 361 Node* u = ctl->fast_out(i); 362 if (u != ctl) { 363 igvn->rehash_node_delayed(u); 364 int nb = u->replace_edge(ctl, phase->C->top(), igvn); 365 --i, imax -= nb; 366 } 367 } 368 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr)); 369 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address")); 370 phase->C->root()->add_req(halt); 371 return this; 372 } 373 374 if (can_reshape && igvn != nullptr && 375 (igvn->_worklist.member(address) || 376 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) { 377 // The address's base and type may change when the address is processed. 378 // Delay this mem node transformation until the address is processed. 379 igvn->_worklist.push(this); 380 return NodeSentinel; // caller will return null 381 } 382 383 // Do NOT remove or optimize the next lines: ensure a new alias index 384 // is allocated for an oop pointer type before Escape Analysis. 385 // Note: C++ will not remove it since the call has side effect. 386 if (t_adr->isa_oopptr()) { 387 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 388 } 389 390 Node* base = nullptr; 391 if (address->is_AddP()) { 392 base = address->in(AddPNode::Base); 393 } 394 if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 395 !t_adr->isa_rawptr()) { 396 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 397 // Skip this node optimization if its address has TOP base. 398 return NodeSentinel; // caller will return null 399 } 400 401 // Avoid independent memory operations 402 Node* old_mem = mem; 403 404 // The code which unhooks non-raw memories from complete (macro-expanded) 405 // initializations was removed. After macro-expansion all stores caught 406 // by Initialize node became raw stores and there is no information 407 // which memory slices they modify. So it is unsafe to move any memory 408 // operation above these stores. Also in most cases hooked non-raw memories 409 // were already unhooked by using information from detect_ptr_independence() 410 // and find_previous_store(). 411 412 if (mem->is_MergeMem()) { 413 MergeMemNode* mmem = mem->as_MergeMem(); 414 const TypePtr *tp = t_adr->is_ptr(); 415 416 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 417 } 418 419 if (mem != old_mem) { 420 set_req_X(MemNode::Memory, mem, phase); 421 if (phase->type(mem) == Type::TOP) return NodeSentinel; 422 return this; 423 } 424 425 // let the subclass continue analyzing... 426 return nullptr; 427 } 428 429 // Helper function for proving some simple control dominations. 430 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 431 // Already assumes that 'dom' is available at 'sub', and that 'sub' 432 // is not a constant (dominated by the method's StartNode). 433 // Used by MemNode::find_previous_store to prove that the 434 // control input of a memory operation predates (dominates) 435 // an allocation it wants to look past. 436 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 437 if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top()) 438 return false; // Conservative answer for dead code 439 440 // Check 'dom'. Skip Proj and CatchProj nodes. 441 dom = dom->find_exact_control(dom); 442 if (dom == nullptr || dom->is_top()) 443 return false; // Conservative answer for dead code 444 445 if (dom == sub) { 446 // For the case when, for example, 'sub' is Initialize and the original 447 // 'dom' is Proj node of the 'sub'. 448 return false; 449 } 450 451 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 452 return true; 453 454 // 'dom' dominates 'sub' if its control edge and control edges 455 // of all its inputs dominate or equal to sub's control edge. 456 457 // Currently 'sub' is either Allocate, Initialize or Start nodes. 458 // Or Region for the check in LoadNode::Ideal(); 459 // 'sub' should have sub->in(0) != nullptr. 460 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 461 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 462 463 // Get control edge of 'sub'. 464 Node* orig_sub = sub; 465 sub = sub->find_exact_control(sub->in(0)); 466 if (sub == nullptr || sub->is_top()) 467 return false; // Conservative answer for dead code 468 469 assert(sub->is_CFG(), "expecting control"); 470 471 if (sub == dom) 472 return true; 473 474 if (sub->is_Start() || sub->is_Root()) 475 return false; 476 477 { 478 // Check all control edges of 'dom'. 479 480 ResourceMark rm; 481 Node_List nlist; 482 Unique_Node_List dom_list; 483 484 dom_list.push(dom); 485 bool only_dominating_controls = false; 486 487 for (uint next = 0; next < dom_list.size(); next++) { 488 Node* n = dom_list.at(next); 489 if (n == orig_sub) 490 return false; // One of dom's inputs dominated by sub. 491 if (!n->is_CFG() && n->pinned()) { 492 // Check only own control edge for pinned non-control nodes. 493 n = n->find_exact_control(n->in(0)); 494 if (n == nullptr || n->is_top()) 495 return false; // Conservative answer for dead code 496 assert(n->is_CFG(), "expecting control"); 497 dom_list.push(n); 498 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 499 only_dominating_controls = true; 500 } else if (n->is_CFG()) { 501 if (n->dominates(sub, nlist)) 502 only_dominating_controls = true; 503 else 504 return false; 505 } else { 506 // First, own control edge. 507 Node* m = n->find_exact_control(n->in(0)); 508 if (m != nullptr) { 509 if (m->is_top()) 510 return false; // Conservative answer for dead code 511 dom_list.push(m); 512 } 513 // Now, the rest of edges. 514 uint cnt = n->req(); 515 for (uint i = 1; i < cnt; i++) { 516 m = n->find_exact_control(n->in(i)); 517 if (m == nullptr || m->is_top()) 518 continue; 519 dom_list.push(m); 520 } 521 } 522 } 523 return only_dominating_controls; 524 } 525 } 526 527 //---------------------detect_ptr_independence--------------------------------- 528 // Used by MemNode::find_previous_store to prove that two base 529 // pointers are never equal. 530 // The pointers are accompanied by their associated allocations, 531 // if any, which have been previously discovered by the caller. 532 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 533 Node* p2, AllocateNode* a2, 534 PhaseTransform* phase) { 535 // Attempt to prove that these two pointers cannot be aliased. 536 // They may both manifestly be allocations, and they should differ. 537 // Or, if they are not both allocations, they can be distinct constants. 538 // Otherwise, one is an allocation and the other a pre-existing value. 539 if (a1 == nullptr && a2 == nullptr) { // neither an allocation 540 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 541 } else if (a1 != nullptr && a2 != nullptr) { // both allocations 542 return (a1 != a2); 543 } else if (a1 != nullptr) { // one allocation a1 544 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 545 return all_controls_dominate(p2, a1); 546 } else { //(a2 != null) // one allocation a2 547 return all_controls_dominate(p1, a2); 548 } 549 return false; 550 } 551 552 553 // Find an arraycopy ac that produces the memory state represented by parameter mem. 554 // Return ac if 555 // (a) can_see_stored_value=true and ac must have set the value for this load or if 556 // (b) can_see_stored_value=false and ac could have set the value for this load or if 557 // (c) can_see_stored_value=false and ac cannot have set the value for this load. 558 // In case (c) change the parameter mem to the memory input of ac to skip it 559 // when searching stored value. 560 // Otherwise return null. 561 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const { 562 ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem); 563 if (ac != nullptr) { 564 Node* ld_addp = in(MemNode::Address); 565 Node* src = ac->in(ArrayCopyNode::Src); 566 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr(); 567 568 // This is a load from a cloned array. The corresponding arraycopy ac must 569 // have set the value for the load and we can return ac but only if the load 570 // is known to be within bounds. This is checked below. 571 if (ary_t != nullptr && ld_addp->is_AddP()) { 572 Node* ld_offs = ld_addp->in(AddPNode::Offset); 573 BasicType ary_elem = ary_t->elem()->array_element_basic_type(); 574 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem); 575 jlong elemsize = type2aelembytes(ary_elem); 576 577 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 578 const TypeInt* sizetype = ary_t->size(); 579 580 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) { 581 // The load is known to be within bounds. It receives its value from ac. 582 return ac; 583 } 584 // The load is known to be out-of-bounds. 585 } 586 // The load could be out-of-bounds. It must not be hoisted but must remain 587 // dependent on the runtime range check. This is achieved by returning null. 588 } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) { 589 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy(); 590 591 if (ac->is_arraycopy_validated() || 592 ac->is_copyof_validated() || 593 ac->is_copyofrange_validated()) { 594 Node* ld_addp = in(MemNode::Address); 595 if (ld_addp->is_AddP()) { 596 Node* ld_base = ld_addp->in(AddPNode::Address); 597 Node* ld_offs = ld_addp->in(AddPNode::Offset); 598 599 Node* dest = ac->in(ArrayCopyNode::Dest); 600 601 if (dest == ld_base) { 602 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t(); 603 assert(!ld_offs_t->empty(), "dead reference should be checked already"); 604 // Take into account vector or unsafe access size 605 jlong ld_size_in_bytes = (jlong)memory_size(); 606 jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1; 607 offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM 608 if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) { 609 return ac; 610 } 611 if (!can_see_stored_value) { 612 mem = ac->in(TypeFunc::Memory); 613 return ac; 614 } 615 } 616 } 617 } 618 } 619 return nullptr; 620 } 621 622 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const { 623 if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore || 624 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) { 625 if (ld_alloc != nullptr) { 626 // Check if there is an array copy for a clone 627 Node* mb = mem->in(0); 628 ArrayCopyNode* ac = nullptr; 629 if (mb->in(0) != nullptr && mb->in(0)->is_Proj() && 630 mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) { 631 ac = mb->in(0)->in(0)->as_ArrayCopy(); 632 } else { 633 // Step over GC barrier when ReduceInitialCardMarks is disabled 634 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 635 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0)); 636 637 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) { 638 ac = control_proj_ac->in(0)->as_ArrayCopy(); 639 } 640 } 641 642 if (ac != nullptr && ac->is_clonebasic()) { 643 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest)); 644 if (alloc != nullptr && alloc == ld_alloc) { 645 return ac; 646 } 647 } 648 } 649 } 650 return nullptr; 651 } 652 653 // The logic for reordering loads and stores uses four steps: 654 // (a) Walk carefully past stores and initializations which we 655 // can prove are independent of this load. 656 // (b) Observe that the next memory state makes an exact match 657 // with self (load or store), and locate the relevant store. 658 // (c) Ensure that, if we were to wire self directly to the store, 659 // the optimizer would fold it up somehow. 660 // (d) Do the rewiring, and return, depending on some other part of 661 // the optimizer to fold up the load. 662 // This routine handles steps (a) and (b). Steps (c) and (d) are 663 // specific to loads and stores, so they are handled by the callers. 664 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 665 // 666 Node* MemNode::find_previous_store(PhaseValues* phase) { 667 Node* ctrl = in(MemNode::Control); 668 Node* adr = in(MemNode::Address); 669 intptr_t offset = 0; 670 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 671 AllocateNode* alloc = AllocateNode::Ideal_allocation(base); 672 673 if (offset == Type::OffsetBot) 674 return nullptr; // cannot unalias unless there are precise offsets 675 676 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr); 677 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 678 679 intptr_t size_in_bytes = memory_size(); 680 681 Node* mem = in(MemNode::Memory); // start searching here... 682 683 int cnt = 50; // Cycle limiter 684 for (;;) { // While we can dance past unrelated stores... 685 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 686 687 Node* prev = mem; 688 if (mem->is_Store()) { 689 Node* st_adr = mem->in(MemNode::Address); 690 intptr_t st_offset = 0; 691 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 692 if (st_base == nullptr) 693 break; // inscrutable pointer 694 695 // For raw accesses it's not enough to prove that constant offsets don't intersect. 696 // We need the bases to be the equal in order for the offset check to make sense. 697 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) { 698 break; 699 } 700 701 if (st_offset != offset && st_offset != Type::OffsetBot) { 702 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize); 703 assert(mem->as_Store()->memory_size() <= MAX_STORE, ""); 704 if (st_offset >= offset + size_in_bytes || 705 st_offset <= offset - MAX_STORE || 706 st_offset <= offset - mem->as_Store()->memory_size()) { 707 // Success: The offsets are provably independent. 708 // (You may ask, why not just test st_offset != offset and be done? 709 // The answer is that stores of different sizes can co-exist 710 // in the same sequence of RawMem effects. We sometimes initialize 711 // a whole 'tile' of array elements with a single jint or jlong.) 712 mem = mem->in(MemNode::Memory); 713 continue; // (a) advance through independent store memory 714 } 715 } 716 if (st_base != base && 717 detect_ptr_independence(base, alloc, 718 st_base, 719 AllocateNode::Ideal_allocation(st_base), 720 phase)) { 721 // Success: The bases are provably independent. 722 mem = mem->in(MemNode::Memory); 723 continue; // (a) advance through independent store memory 724 } 725 726 // (b) At this point, if the bases or offsets do not agree, we lose, 727 // since we have not managed to prove 'this' and 'mem' independent. 728 if (st_base == base && st_offset == offset) { 729 return mem; // let caller handle steps (c), (d) 730 } 731 732 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 733 InitializeNode* st_init = mem->in(0)->as_Initialize(); 734 AllocateNode* st_alloc = st_init->allocation(); 735 if (st_alloc == nullptr) 736 break; // something degenerated 737 bool known_identical = false; 738 bool known_independent = false; 739 if (alloc == st_alloc) 740 known_identical = true; 741 else if (alloc != nullptr) 742 known_independent = true; 743 else if (all_controls_dominate(this, st_alloc)) 744 known_independent = true; 745 746 if (known_independent) { 747 // The bases are provably independent: Either they are 748 // manifestly distinct allocations, or else the control 749 // of this load dominates the store's allocation. 750 int alias_idx = phase->C->get_alias_index(adr_type()); 751 if (alias_idx == Compile::AliasIdxRaw) { 752 mem = st_alloc->in(TypeFunc::Memory); 753 } else { 754 mem = st_init->memory(alias_idx); 755 } 756 continue; // (a) advance through independent store memory 757 } 758 759 // (b) at this point, if we are not looking at a store initializing 760 // the same allocation we are loading from, we lose. 761 if (known_identical) { 762 // From caller, can_see_stored_value will consult find_captured_store. 763 return mem; // let caller handle steps (c), (d) 764 } 765 766 } else if (find_previous_arraycopy(phase, alloc, mem, false) != nullptr) { 767 if (prev != mem) { 768 // Found an arraycopy but it doesn't affect that load 769 continue; 770 } 771 // Found an arraycopy that may affect that load 772 return mem; 773 } else if (addr_t != nullptr && addr_t->is_known_instance_field()) { 774 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 775 if (mem->is_Proj() && mem->in(0)->is_Call()) { 776 // ArrayCopyNodes processed here as well. 777 CallNode *call = mem->in(0)->as_Call(); 778 if (!call->may_modify(addr_t, phase)) { 779 mem = call->in(TypeFunc::Memory); 780 continue; // (a) advance through independent call memory 781 } 782 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 783 ArrayCopyNode* ac = nullptr; 784 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) { 785 break; 786 } 787 mem = mem->in(0)->in(TypeFunc::Memory); 788 continue; // (a) advance through independent MemBar memory 789 } else if (mem->is_ClearArray()) { 790 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 791 // (the call updated 'mem' value) 792 continue; // (a) advance through independent allocation memory 793 } else { 794 // Can not bypass initialization of the instance 795 // we are looking for. 796 return mem; 797 } 798 } else if (mem->is_MergeMem()) { 799 int alias_idx = phase->C->get_alias_index(adr_type()); 800 mem = mem->as_MergeMem()->memory_at(alias_idx); 801 continue; // (a) advance through independent MergeMem memory 802 } 803 } 804 805 // Unless there is an explicit 'continue', we must bail out here, 806 // because 'mem' is an inscrutable memory state (e.g., a call). 807 break; 808 } 809 810 return nullptr; // bail out 811 } 812 813 //----------------------calculate_adr_type------------------------------------- 814 // Helper function. Notices when the given type of address hits top or bottom. 815 // Also, asserts a cross-check of the type against the expected address type. 816 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 817 if (t == Type::TOP) return nullptr; // does not touch memory any more? 818 #ifdef ASSERT 819 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = nullptr; 820 #endif 821 const TypePtr* tp = t->isa_ptr(); 822 if (tp == nullptr) { 823 assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 824 return TypePtr::BOTTOM; // touches lots of memory 825 } else { 826 #ifdef ASSERT 827 // %%%% [phh] We don't check the alias index if cross_check is 828 // TypeRawPtr::BOTTOM. Needs to be investigated. 829 if (cross_check != nullptr && 830 cross_check != TypePtr::BOTTOM && 831 cross_check != TypeRawPtr::BOTTOM) { 832 // Recheck the alias index, to see if it has changed (due to a bug). 833 Compile* C = Compile::current(); 834 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 835 "must stay in the original alias category"); 836 // The type of the address must be contained in the adr_type, 837 // disregarding "null"-ness. 838 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 839 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 840 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 841 "real address must not escape from expected memory type"); 842 } 843 #endif 844 return tp; 845 } 846 } 847 848 uint8_t MemNode::barrier_data(const Node* n) { 849 if (n->is_LoadStore()) { 850 return n->as_LoadStore()->barrier_data(); 851 } else if (n->is_Mem()) { 852 return n->as_Mem()->barrier_data(); 853 } 854 return 0; 855 } 856 857 //============================================================================= 858 // Should LoadNode::Ideal() attempt to remove control edges? 859 bool LoadNode::can_remove_control() const { 860 return !has_pinned_control_dependency(); 861 } 862 uint LoadNode::size_of() const { return sizeof(*this); } 863 bool LoadNode::cmp(const Node &n) const { 864 LoadNode& load = (LoadNode &)n; 865 return !Type::cmp(_type, load._type) && 866 _control_dependency == load._control_dependency && 867 _mo == load._mo; 868 } 869 const Type *LoadNode::bottom_type() const { return _type; } 870 uint LoadNode::ideal_reg() const { 871 return _type->ideal_reg(); 872 } 873 874 #ifndef PRODUCT 875 void LoadNode::dump_spec(outputStream *st) const { 876 MemNode::dump_spec(st); 877 if( !Verbose && !WizardMode ) { 878 // standard dump does this in Verbose and WizardMode 879 st->print(" #"); _type->dump_on(st); 880 } 881 if (!depends_only_on_test()) { 882 st->print(" (does not depend only on test, "); 883 if (control_dependency() == UnknownControl) { 884 st->print("unknown control"); 885 } else if (control_dependency() == Pinned) { 886 st->print("pinned"); 887 } else if (adr_type() == TypeRawPtr::BOTTOM) { 888 st->print("raw access"); 889 } else { 890 st->print("unknown reason"); 891 } 892 st->print(")"); 893 } 894 } 895 #endif 896 897 #ifdef ASSERT 898 //----------------------------is_immutable_value------------------------------- 899 // Helper function to allow a raw load without control edge for some cases 900 bool LoadNode::is_immutable_value(Node* adr) { 901 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 902 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) { 903 904 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1); 905 int offsets[] = { 906 in_bytes(JavaThread::osthread_offset()), 907 in_bytes(JavaThread::threadObj_offset()), 908 in_bytes(JavaThread::vthread_offset()), 909 in_bytes(JavaThread::scopedValueCache_offset()), 910 }; 911 912 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) { 913 if (offset == offsets[i]) { 914 return true; 915 } 916 } 917 } 918 919 return false; 920 } 921 #endif 922 923 //----------------------------LoadNode::make----------------------------------- 924 // Polymorphic factory method: 925 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo, 926 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) { 927 Compile* C = gvn.C; 928 929 // sanity check the alias category against the created node type 930 assert(!(adr_type->isa_oopptr() && 931 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 932 "use LoadKlassNode instead"); 933 assert(!(adr_type->isa_aryptr() && 934 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 935 "use LoadRangeNode instead"); 936 // Check control edge of raw loads 937 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 938 // oop will be recorded in oop map if load crosses safepoint 939 rt->isa_oopptr() || is_immutable_value(adr), 940 "raw memory operations should have control edge"); 941 LoadNode* load = nullptr; 942 switch (bt) { 943 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 944 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 945 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 946 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 947 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 948 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break; 949 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 950 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break; 951 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break; 952 case T_OBJECT: 953 #ifdef _LP64 954 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 955 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency); 956 } else 957 #endif 958 { 959 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 960 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 961 } 962 break; 963 default: 964 ShouldNotReachHere(); 965 break; 966 } 967 assert(load != nullptr, "LoadNode should have been created"); 968 if (unaligned) { 969 load->set_unaligned_access(); 970 } 971 if (mismatched) { 972 load->set_mismatched_access(); 973 } 974 if (unsafe) { 975 load->set_unsafe_access(); 976 } 977 load->set_barrier_data(barrier_data); 978 if (load->Opcode() == Op_LoadN) { 979 Node* ld = gvn.transform(load); 980 return new DecodeNNode(ld, ld->bottom_type()->make_ptr()); 981 } 982 983 return load; 984 } 985 986 //------------------------------hash------------------------------------------- 987 uint LoadNode::hash() const { 988 // unroll addition of interesting fields 989 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 990 } 991 992 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 993 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) { 994 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile(); 995 bool is_stable_ary = FoldStableValues && 996 (tp != nullptr) && (tp->isa_aryptr() != nullptr) && 997 tp->isa_aryptr()->is_stable(); 998 999 return (eliminate_boxing && non_volatile) || is_stable_ary || tp->is_inlinetypeptr(); 1000 } 1001 1002 return false; 1003 } 1004 1005 LoadNode* LoadNode::pin_array_access_node() const { 1006 const TypePtr* adr_type = this->adr_type(); 1007 if (adr_type != nullptr && adr_type->isa_aryptr()) { 1008 return clone_pinned(); 1009 } 1010 return nullptr; 1011 } 1012 1013 // Is the value loaded previously stored by an arraycopy? If so return 1014 // a load node that reads from the source array so we may be able to 1015 // optimize out the ArrayCopy node later. 1016 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const { 1017 Node* ld_adr = in(MemNode::Address); 1018 intptr_t ld_off = 0; 1019 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 1020 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 1021 if (ac != nullptr) { 1022 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 1023 1024 Node* mem = ac->in(TypeFunc::Memory); 1025 Node* ctl = ac->in(0); 1026 Node* src = ac->in(ArrayCopyNode::Src); 1027 1028 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) { 1029 return nullptr; 1030 } 1031 1032 // load depends on the tests that validate the arraycopy 1033 LoadNode* ld = clone_pinned(); 1034 Node* addp = in(MemNode::Address)->clone(); 1035 if (ac->as_ArrayCopy()->is_clonebasic()) { 1036 assert(ld_alloc != nullptr, "need an alloc"); 1037 assert(addp->is_AddP(), "address must be addp"); 1038 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1039 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern"); 1040 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern"); 1041 addp->set_req(AddPNode::Base, src); 1042 addp->set_req(AddPNode::Address, src); 1043 } else { 1044 assert(ac->as_ArrayCopy()->is_arraycopy_validated() || 1045 ac->as_ArrayCopy()->is_copyof_validated() || 1046 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases"); 1047 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 1048 addp->set_req(AddPNode::Base, src); 1049 addp->set_req(AddPNode::Address, src); 1050 1051 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 1052 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type(); 1053 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT; 1054 1055 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 1056 uint shift = ary_t->is_flat() ? ary_t->flat_log_elem_size() : exact_log2(type2aelembytes(ary_elem)); 1057 1058 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 1059 #ifdef _LP64 1060 diff = phase->transform(new ConvI2LNode(diff)); 1061 #endif 1062 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 1063 1064 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 1065 addp->set_req(AddPNode::Offset, offset); 1066 } 1067 addp = phase->transform(addp); 1068 #ifdef ASSERT 1069 const TypePtr* adr_type = phase->type(addp)->is_ptr(); 1070 ld->_adr_type = adr_type; 1071 #endif 1072 ld->set_req(MemNode::Address, addp); 1073 ld->set_req(0, ctl); 1074 ld->set_req(MemNode::Memory, mem); 1075 return ld; 1076 } 1077 return nullptr; 1078 } 1079 1080 1081 //---------------------------can_see_stored_value------------------------------ 1082 // This routine exists to make sure this set of tests is done the same 1083 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 1084 // will change the graph shape in a way which makes memory alive twice at the 1085 // same time (uses the Oracle model of aliasing), then some 1086 // LoadXNode::Identity will fold things back to the equivalence-class model 1087 // of aliasing. 1088 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const { 1089 Node* ld_adr = in(MemNode::Address); 1090 intptr_t ld_off = 0; 1091 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off); 1092 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base); 1093 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 1094 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr; 1095 // This is more general than load from boxing objects. 1096 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 1097 uint alias_idx = atp->index(); 1098 Node* result = nullptr; 1099 Node* current = st; 1100 // Skip through chains of MemBarNodes checking the MergeMems for 1101 // new states for the slice of this load. Stop once any other 1102 // kind of node is encountered. Loads from final memory can skip 1103 // through any kind of MemBar but normal loads shouldn't skip 1104 // through MemBarAcquire since the could allow them to move out of 1105 // a synchronized region. It is not safe to step over MemBarCPUOrder, 1106 // because alias info above them may be inaccurate (e.g., due to 1107 // mixed/mismatched unsafe accesses). 1108 bool is_final_mem = !atp->is_rewritable(); 1109 while (current->is_Proj()) { 1110 int opc = current->in(0)->Opcode(); 1111 if ((is_final_mem && (opc == Op_MemBarAcquire || 1112 opc == Op_MemBarAcquireLock || 1113 opc == Op_LoadFence)) || 1114 opc == Op_MemBarRelease || 1115 opc == Op_StoreFence || 1116 opc == Op_MemBarReleaseLock || 1117 opc == Op_MemBarStoreStore || 1118 opc == Op_StoreStoreFence) { 1119 Node* mem = current->in(0)->in(TypeFunc::Memory); 1120 if (mem->is_MergeMem()) { 1121 MergeMemNode* merge = mem->as_MergeMem(); 1122 Node* new_st = merge->memory_at(alias_idx); 1123 if (new_st == merge->base_memory()) { 1124 // Keep searching 1125 current = new_st; 1126 continue; 1127 } 1128 // Save the new memory state for the slice and fall through 1129 // to exit. 1130 result = new_st; 1131 } 1132 } 1133 break; 1134 } 1135 if (result != nullptr) { 1136 st = result; 1137 } 1138 } 1139 1140 // Loop around twice in the case Load -> Initialize -> Store. 1141 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 1142 for (int trip = 0; trip <= 1; trip++) { 1143 1144 if (st->is_Store()) { 1145 Node* st_adr = st->in(MemNode::Address); 1146 if (st_adr != ld_adr) { 1147 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers). 1148 intptr_t st_off = 0; 1149 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off); 1150 if (ld_base == nullptr) return nullptr; 1151 if (st_base == nullptr) return nullptr; 1152 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr; 1153 if (ld_off != st_off) return nullptr; 1154 if (ld_off == Type::OffsetBot) return nullptr; 1155 // Same base, same offset. 1156 // Possible improvement for arrays: check index value instead of absolute offset. 1157 1158 // At this point we have proven something like this setup: 1159 // B = << base >> 1160 // L = LoadQ(AddP(Check/CastPP(B), #Off)) 1161 // S = StoreQ(AddP( B , #Off), V) 1162 // (Actually, we haven't yet proven the Q's are the same.) 1163 // In other words, we are loading from a casted version of 1164 // the same pointer-and-offset that we stored to. 1165 // Casted version may carry a dependency and it is respected. 1166 // Thus, we are able to replace L by V. 1167 } 1168 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1169 if (store_Opcode() != st->Opcode()) { 1170 return nullptr; 1171 } 1172 // LoadVector/StoreVector needs additional check to ensure the types match. 1173 if (st->is_StoreVector()) { 1174 const TypeVect* in_vt = st->as_StoreVector()->vect_type(); 1175 const TypeVect* out_vt = is_Load() ? as_LoadVector()->vect_type() : as_StoreVector()->vect_type(); 1176 if (in_vt != out_vt) { 1177 return nullptr; 1178 } 1179 } 1180 return st->in(MemNode::ValueIn); 1181 } 1182 1183 // A load from a freshly-created object always returns zero. 1184 // (This can happen after LoadNode::Ideal resets the load's memory input 1185 // to find_captured_store, which returned InitializeNode::zero_memory.) 1186 if (st->is_Proj() && st->in(0)->is_Allocate() && 1187 (st->in(0) == ld_alloc) && 1188 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1189 // return a zero value for the load's basic type 1190 // (This is one of the few places where a generic PhaseTransform 1191 // can create new nodes. Think of it as lazily manifesting 1192 // virtually pre-existing constants.) 1193 Node* default_value = ld_alloc->in(AllocateNode::DefaultValue); 1194 if (default_value != nullptr) { 1195 return default_value; 1196 } 1197 assert(ld_alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null"); 1198 if (memory_type() != T_VOID) { 1199 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) { 1200 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an 1201 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done 1202 // by the ArrayCopyNode. 1203 return phase->zerocon(memory_type()); 1204 } 1205 } else { 1206 // TODO: materialize all-zero vector constant 1207 assert(!isa_Load() || as_Load()->type()->isa_vect(), ""); 1208 } 1209 } 1210 1211 // A load from an initialization barrier can match a captured store. 1212 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1213 InitializeNode* init = st->in(0)->as_Initialize(); 1214 AllocateNode* alloc = init->allocation(); 1215 if ((alloc != nullptr) && (alloc == ld_alloc)) { 1216 // examine a captured store value 1217 st = init->find_captured_store(ld_off, memory_size(), phase); 1218 if (st != nullptr) { 1219 continue; // take one more trip around 1220 } 1221 } 1222 } 1223 1224 // Load boxed value from result of valueOf() call is input parameter. 1225 if (this->is_Load() && ld_adr->is_AddP() && 1226 (tp != nullptr) && tp->is_ptr_to_boxed_value()) { 1227 intptr_t ignore = 0; 1228 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1229 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1230 base = bs->step_over_gc_barrier(base); 1231 if (base != nullptr && base->is_Proj() && 1232 base->as_Proj()->_con == TypeFunc::Parms && 1233 base->in(0)->is_CallStaticJava() && 1234 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1235 return base->in(0)->in(TypeFunc::Parms); 1236 } 1237 } 1238 1239 break; 1240 } 1241 1242 return nullptr; 1243 } 1244 1245 //----------------------is_instance_field_load_with_local_phi------------------ 1246 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1247 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1248 in(Address)->is_AddP() ) { 1249 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1250 // Only instances and boxed values. 1251 if( t_oop != nullptr && 1252 (t_oop->is_ptr_to_boxed_value() || 1253 t_oop->is_known_instance_field()) && 1254 t_oop->offset() != Type::OffsetBot && 1255 t_oop->offset() != Type::OffsetTop) { 1256 return true; 1257 } 1258 } 1259 return false; 1260 } 1261 1262 //------------------------------Identity--------------------------------------- 1263 // Loads are identity if previous store is to same address 1264 Node* LoadNode::Identity(PhaseGVN* phase) { 1265 // Loading from an InlineType? The InlineType has the values of 1266 // all fields as input. Look for the field with matching offset. 1267 Node* addr = in(Address); 1268 intptr_t offset; 1269 Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset); 1270 if (base != nullptr && base->is_InlineType() && offset > oopDesc::klass_offset_in_bytes()) { 1271 Node* value = base->as_InlineType()->field_value_by_offset((int)offset, true); 1272 if (value != nullptr) { 1273 if (Opcode() == Op_LoadN) { 1274 // Encode oop value if we are loading a narrow oop 1275 assert(!phase->type(value)->isa_narrowoop(), "should already be decoded"); 1276 value = phase->transform(new EncodePNode(value, bottom_type())); 1277 } 1278 return value; 1279 } 1280 } 1281 1282 // If the previous store-maker is the right kind of Store, and the store is 1283 // to the same address, then we are equal to the value stored. 1284 Node* mem = in(Memory); 1285 Node* value = can_see_stored_value(mem, phase); 1286 if( value ) { 1287 // byte, short & char stores truncate naturally. 1288 // A load has to load the truncated value which requires 1289 // some sort of masking operation and that requires an 1290 // Ideal call instead of an Identity call. 1291 if (memory_size() < BytesPerInt) { 1292 // If the input to the store does not fit with the load's result type, 1293 // it must be truncated via an Ideal call. 1294 if (!phase->type(value)->higher_equal(phase->type(this))) 1295 return this; 1296 } 1297 // (This works even when value is a Con, but LoadNode::Value 1298 // usually runs first, producing the singleton type of the Con.) 1299 if (!has_pinned_control_dependency() || value->is_Con()) { 1300 return value; 1301 } else { 1302 return this; 1303 } 1304 } 1305 1306 if (has_pinned_control_dependency()) { 1307 return this; 1308 } 1309 // Search for an existing data phi which was generated before for the same 1310 // instance's field to avoid infinite generation of phis in a loop. 1311 Node *region = mem->in(0); 1312 if (is_instance_field_load_with_local_phi(region)) { 1313 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1314 int this_index = phase->C->get_alias_index(addr_t); 1315 int this_offset = addr_t->offset(); 1316 int this_iid = addr_t->instance_id(); 1317 if (!addr_t->is_known_instance() && 1318 addr_t->is_ptr_to_boxed_value()) { 1319 // Use _idx of address base (could be Phi node) for boxed values. 1320 intptr_t ignore = 0; 1321 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1322 if (base == nullptr) { 1323 return this; 1324 } 1325 this_iid = base->_idx; 1326 } 1327 const Type* this_type = bottom_type(); 1328 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1329 Node* phi = region->fast_out(i); 1330 if (phi->is_Phi() && phi != mem && 1331 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1332 return phi; 1333 } 1334 } 1335 } 1336 1337 return this; 1338 } 1339 1340 // Construct an equivalent unsigned load. 1341 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) { 1342 BasicType bt = T_ILLEGAL; 1343 const Type* rt = nullptr; 1344 switch (Opcode()) { 1345 case Op_LoadUB: return this; 1346 case Op_LoadUS: return this; 1347 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break; 1348 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break; 1349 default: 1350 assert(false, "no unsigned variant: %s", Name()); 1351 return nullptr; 1352 } 1353 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1354 raw_adr_type(), rt, bt, _mo, _control_dependency, 1355 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access()); 1356 } 1357 1358 // Construct an equivalent signed load. 1359 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) { 1360 BasicType bt = T_ILLEGAL; 1361 const Type* rt = nullptr; 1362 switch (Opcode()) { 1363 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break; 1364 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break; 1365 case Op_LoadB: // fall through 1366 case Op_LoadS: // fall through 1367 case Op_LoadI: // fall through 1368 case Op_LoadL: return this; 1369 default: 1370 assert(false, "no signed variant: %s", Name()); 1371 return nullptr; 1372 } 1373 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1374 raw_adr_type(), rt, bt, _mo, _control_dependency, 1375 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access()); 1376 } 1377 1378 bool LoadNode::has_reinterpret_variant(const Type* rt) { 1379 BasicType bt = rt->basic_type(); 1380 switch (Opcode()) { 1381 case Op_LoadI: return (bt == T_FLOAT); 1382 case Op_LoadL: return (bt == T_DOUBLE); 1383 case Op_LoadF: return (bt == T_INT); 1384 case Op_LoadD: return (bt == T_LONG); 1385 1386 default: return false; 1387 } 1388 } 1389 1390 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) { 1391 BasicType bt = rt->basic_type(); 1392 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt)); 1393 bool is_mismatched = is_mismatched_access(); 1394 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr(); 1395 if (raw_type == nullptr) { 1396 is_mismatched = true; // conservatively match all non-raw accesses as mismatched 1397 } 1398 const int op = Opcode(); 1399 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) || 1400 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access()); 1401 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1402 raw_adr_type(), rt, bt, _mo, _control_dependency, 1403 require_atomic_access, is_unaligned_access(), is_mismatched); 1404 } 1405 1406 bool StoreNode::has_reinterpret_variant(const Type* vt) { 1407 BasicType bt = vt->basic_type(); 1408 switch (Opcode()) { 1409 case Op_StoreI: return (bt == T_FLOAT); 1410 case Op_StoreL: return (bt == T_DOUBLE); 1411 case Op_StoreF: return (bt == T_INT); 1412 case Op_StoreD: return (bt == T_LONG); 1413 1414 default: return false; 1415 } 1416 } 1417 1418 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) { 1419 BasicType bt = vt->basic_type(); 1420 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt)); 1421 const int op = Opcode(); 1422 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) || 1423 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access()); 1424 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1425 raw_adr_type(), val, bt, _mo, require_atomic_access); 1426 1427 bool is_mismatched = is_mismatched_access(); 1428 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr(); 1429 if (raw_type == nullptr) { 1430 is_mismatched = true; // conservatively match all non-raw accesses as mismatched 1431 } 1432 if (is_mismatched) { 1433 st->set_mismatched_access(); 1434 } 1435 return st; 1436 } 1437 1438 // We're loading from an object which has autobox behaviour. 1439 // If this object is result of a valueOf call we'll have a phi 1440 // merging a newly allocated object and a load from the cache. 1441 // We want to replace this load with the original incoming 1442 // argument to the valueOf call. 1443 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) { 1444 assert(igvn->C->eliminate_boxing(), "sanity"); 1445 intptr_t ignore = 0; 1446 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore); 1447 if ((base == nullptr) || base->is_Phi()) { 1448 // Push the loads from the phi that comes from valueOf up 1449 // through it to allow elimination of the loads and the recovery 1450 // of the original value. It is done in split_through_phi(). 1451 return nullptr; 1452 } else if (base->is_Load() || 1453 (base->is_DecodeN() && base->in(1)->is_Load())) { 1454 // Eliminate the load of boxed value for integer types from the cache 1455 // array by deriving the value from the index into the array. 1456 // Capture the offset of the load and then reverse the computation. 1457 1458 // Get LoadN node which loads a boxing object from 'cache' array. 1459 if (base->is_DecodeN()) { 1460 base = base->in(1); 1461 } 1462 if (!base->in(Address)->is_AddP()) { 1463 return nullptr; // Complex address 1464 } 1465 AddPNode* address = base->in(Address)->as_AddP(); 1466 Node* cache_base = address->in(AddPNode::Base); 1467 if ((cache_base != nullptr) && cache_base->is_DecodeN()) { 1468 // Get ConP node which is static 'cache' field. 1469 cache_base = cache_base->in(1); 1470 } 1471 if ((cache_base != nullptr) && cache_base->is_Con()) { 1472 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1473 if ((base_type != nullptr) && base_type->is_autobox_cache()) { 1474 Node* elements[4]; 1475 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1476 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1477 if (count > 0 && elements[0]->is_Con() && 1478 (count == 1 || 1479 (count == 2 && elements[1]->Opcode() == Op_LShiftX && 1480 elements[1]->in(2) == igvn->intcon(shift)))) { 1481 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1482 // Fetch the box object cache[0] at the base of the array and get its value 1483 ciInstance* box = array->obj_at(0)->as_instance(); 1484 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1485 assert(ik->is_box_klass(), "sanity"); 1486 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1487 if (ik->nof_nonstatic_fields() == 1) { 1488 // This should be true nonstatic_field_at requires calling 1489 // nof_nonstatic_fields so check it anyway 1490 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1491 BasicType bt = c.basic_type(); 1492 // Only integer types have boxing cache. 1493 assert(bt == T_BOOLEAN || bt == T_CHAR || 1494 bt == T_BYTE || bt == T_SHORT || 1495 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt)); 1496 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1497 if (cache_low != (int)cache_low) { 1498 return nullptr; // should not happen since cache is array indexed by value 1499 } 1500 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1501 if (offset != (int)offset) { 1502 return nullptr; // should not happen since cache is array indexed by value 1503 } 1504 // Add up all the offsets making of the address of the load 1505 Node* result = elements[0]; 1506 for (int i = 1; i < count; i++) { 1507 result = igvn->transform(new AddXNode(result, elements[i])); 1508 } 1509 // Remove the constant offset from the address and then 1510 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset))); 1511 // remove the scaling of the offset to recover the original index. 1512 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) { 1513 // Peel the shift off directly but wrap it in a dummy node 1514 // since Ideal can't return existing nodes 1515 igvn->_worklist.push(result); // remove dead node later 1516 result = new RShiftXNode(result->in(1), igvn->intcon(0)); 1517 } else if (result->is_Add() && result->in(2)->is_Con() && 1518 result->in(1)->Opcode() == Op_LShiftX && 1519 result->in(1)->in(2) == igvn->intcon(shift)) { 1520 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1521 // but for boxing cache access we know that X<<Z will not overflow 1522 // (there is range check) so we do this optimizatrion by hand here. 1523 igvn->_worklist.push(result); // remove dead node later 1524 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift)); 1525 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con)); 1526 } else { 1527 result = new RShiftXNode(result, igvn->intcon(shift)); 1528 } 1529 #ifdef _LP64 1530 if (bt != T_LONG) { 1531 result = new ConvL2INode(igvn->transform(result)); 1532 } 1533 #else 1534 if (bt == T_LONG) { 1535 result = new ConvI2LNode(igvn->transform(result)); 1536 } 1537 #endif 1538 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1539 // Need to preserve unboxing load type if it is unsigned. 1540 switch(this->Opcode()) { 1541 case Op_LoadUB: 1542 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF)); 1543 break; 1544 case Op_LoadUS: 1545 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF)); 1546 break; 1547 } 1548 return result; 1549 } 1550 } 1551 } 1552 } 1553 } 1554 return nullptr; 1555 } 1556 1557 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1558 Node* region = phi->in(0); 1559 if (region == nullptr) { 1560 return false; // Wait stable graph 1561 } 1562 uint cnt = phi->req(); 1563 for (uint i = 1; i < cnt; i++) { 1564 Node* rc = region->in(i); 1565 if (rc == nullptr || phase->type(rc) == Type::TOP) 1566 return false; // Wait stable graph 1567 Node* in = phi->in(i); 1568 if (in == nullptr || phase->type(in) == Type::TOP) 1569 return false; // Wait stable graph 1570 } 1571 return true; 1572 } 1573 1574 //------------------------------split_through_phi------------------------------ 1575 // Check whether a call to 'split_through_phi' would split this load through the 1576 // Phi *base*. This method is essentially a copy of the validations performed 1577 // by 'split_through_phi'. The first use of this method was in EA code as part 1578 // of simplification of allocation merges. 1579 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) { 1580 Node* mem = in(Memory); 1581 Node* address = in(Address); 1582 intptr_t ignore = 0; 1583 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1584 bool base_is_phi = (base != nullptr) && base->is_Phi(); 1585 1586 if (req() > 3 || !base_is_phi) { 1587 return false; 1588 } 1589 1590 if (!mem->is_Phi()) { 1591 if (!MemNode::all_controls_dominate(mem, base->in(0))) 1592 return false; 1593 } else if (base->in(0) != mem->in(0)) { 1594 if (!MemNode::all_controls_dominate(mem, base->in(0))) { 1595 return false; 1596 } 1597 } 1598 1599 return true; 1600 } 1601 1602 //------------------------------split_through_phi------------------------------ 1603 // Split instance or boxed field load through Phi. 1604 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) { 1605 if (req() > 3) { 1606 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs"); 1607 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account 1608 return nullptr; 1609 } 1610 Node* mem = in(Memory); 1611 Node* address = in(Address); 1612 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1613 1614 assert((t_oop != nullptr) && 1615 (ignore_missing_instance_id || 1616 t_oop->is_known_instance_field() || 1617 t_oop->is_ptr_to_boxed_value()), "invalid conditions"); 1618 1619 Compile* C = phase->C; 1620 intptr_t ignore = 0; 1621 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1622 bool base_is_phi = (base != nullptr) && base->is_Phi(); 1623 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1624 (base != nullptr) && (base == address->in(AddPNode::Base)) && 1625 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1626 1627 if (!((mem->is_Phi() || base_is_phi) && 1628 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) { 1629 return nullptr; // Neither memory or base are Phi 1630 } 1631 1632 if (mem->is_Phi()) { 1633 if (!stable_phi(mem->as_Phi(), phase)) { 1634 return nullptr; // Wait stable graph 1635 } 1636 uint cnt = mem->req(); 1637 // Check for loop invariant memory. 1638 if (cnt == 3) { 1639 for (uint i = 1; i < cnt; i++) { 1640 Node* in = mem->in(i); 1641 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1642 if (m == mem) { 1643 if (i == 1) { 1644 // if the first edge was a loop, check second edge too. 1645 // If both are replaceable - we are in an infinite loop 1646 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase); 1647 if (n == mem) { 1648 break; 1649 } 1650 } 1651 set_req(Memory, mem->in(cnt - i)); 1652 return this; // made change 1653 } 1654 } 1655 } 1656 } 1657 if (base_is_phi) { 1658 if (!stable_phi(base->as_Phi(), phase)) { 1659 return nullptr; // Wait stable graph 1660 } 1661 uint cnt = base->req(); 1662 // Check for loop invariant memory. 1663 if (cnt == 3) { 1664 for (uint i = 1; i < cnt; i++) { 1665 if (base->in(i) == base) { 1666 return nullptr; // Wait stable graph 1667 } 1668 } 1669 } 1670 } 1671 1672 // Split through Phi (see original code in loopopts.cpp). 1673 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type"); 1674 1675 // Do nothing here if Identity will find a value 1676 // (to avoid infinite chain of value phis generation). 1677 if (this != Identity(phase)) { 1678 return nullptr; 1679 } 1680 1681 // Select Region to split through. 1682 Node* region; 1683 if (!base_is_phi) { 1684 assert(mem->is_Phi(), "sanity"); 1685 region = mem->in(0); 1686 // Skip if the region dominates some control edge of the address. 1687 if (!MemNode::all_controls_dominate(address, region)) 1688 return nullptr; 1689 } else if (!mem->is_Phi()) { 1690 assert(base_is_phi, "sanity"); 1691 region = base->in(0); 1692 // Skip if the region dominates some control edge of the memory. 1693 if (!MemNode::all_controls_dominate(mem, region)) 1694 return nullptr; 1695 } else if (base->in(0) != mem->in(0)) { 1696 assert(base_is_phi && mem->is_Phi(), "sanity"); 1697 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1698 region = base->in(0); 1699 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1700 region = mem->in(0); 1701 } else { 1702 return nullptr; // complex graph 1703 } 1704 } else { 1705 assert(base->in(0) == mem->in(0), "sanity"); 1706 region = mem->in(0); 1707 } 1708 1709 Node* phi = nullptr; 1710 const Type* this_type = this->bottom_type(); 1711 PhaseIterGVN* igvn = phase->is_IterGVN(); 1712 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) { 1713 int this_index = C->get_alias_index(t_oop); 1714 int this_offset = t_oop->offset(); 1715 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx; 1716 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset); 1717 } else if (ignore_missing_instance_id) { 1718 phi = new PhiNode(region, this_type, nullptr, mem->_idx); 1719 } else { 1720 return nullptr; 1721 } 1722 1723 for (uint i = 1; i < region->req(); i++) { 1724 Node* x; 1725 Node* the_clone = nullptr; 1726 Node* in = region->in(i); 1727 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl && 1728 in != nullptr && in->is_OuterStripMinedLoop()) { 1729 // No node should go in the outer strip mined loop 1730 in = in->in(LoopNode::EntryControl); 1731 } 1732 if (in == nullptr || in == C->top()) { 1733 x = C->top(); // Dead path? Use a dead data op 1734 } else { 1735 x = this->clone(); // Else clone up the data op 1736 the_clone = x; // Remember for possible deletion. 1737 // Alter data node to use pre-phi inputs 1738 if (this->in(0) == region) { 1739 x->set_req(0, in); 1740 } else { 1741 x->set_req(0, nullptr); 1742 } 1743 if (mem->is_Phi() && (mem->in(0) == region)) { 1744 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1745 } 1746 if (address->is_Phi() && address->in(0) == region) { 1747 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1748 } 1749 if (base_is_phi && (base->in(0) == region)) { 1750 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1751 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1752 x->set_req(Address, adr_x); 1753 } 1754 } 1755 // Check for a 'win' on some paths 1756 const Type *t = x->Value(igvn); 1757 1758 bool singleton = t->singleton(); 1759 1760 // See comments in PhaseIdealLoop::split_thru_phi(). 1761 if (singleton && t == Type::TOP) { 1762 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1763 } 1764 1765 if (singleton) { 1766 x = igvn->makecon(t); 1767 } else { 1768 // We now call Identity to try to simplify the cloned node. 1769 // Note that some Identity methods call phase->type(this). 1770 // Make sure that the type array is big enough for 1771 // our new node, even though we may throw the node away. 1772 // (This tweaking with igvn only works because x is a new node.) 1773 igvn->set_type(x, t); 1774 // If x is a TypeNode, capture any more-precise type permanently into Node 1775 // otherwise it will be not updated during igvn->transform since 1776 // igvn->type(x) is set to x->Value() already. 1777 x->raise_bottom_type(t); 1778 Node* y = x->Identity(igvn); 1779 if (y != x) { 1780 x = y; 1781 } else { 1782 y = igvn->hash_find_insert(x); 1783 if (y) { 1784 x = y; 1785 } else { 1786 // Else x is a new node we are keeping 1787 // We do not need register_new_node_with_optimizer 1788 // because set_type has already been called. 1789 igvn->_worklist.push(x); 1790 } 1791 } 1792 } 1793 if (x != the_clone && the_clone != nullptr) { 1794 igvn->remove_dead_node(the_clone); 1795 } 1796 phi->set_req(i, x); 1797 } 1798 // Record Phi 1799 igvn->register_new_node_with_optimizer(phi); 1800 return phi; 1801 } 1802 1803 AllocateNode* LoadNode::is_new_object_mark_load() const { 1804 if (Opcode() == Op_LoadX) { 1805 Node* address = in(MemNode::Address); 1806 AllocateNode* alloc = AllocateNode::Ideal_allocation(address); 1807 Node* mem = in(MemNode::Memory); 1808 if (alloc != nullptr && mem->is_Proj() && 1809 mem->in(0) != nullptr && 1810 mem->in(0) == alloc->initialization() && 1811 alloc->initialization()->proj_out_or_null(0) != nullptr) { 1812 return alloc; 1813 } 1814 } 1815 return nullptr; 1816 } 1817 1818 1819 //------------------------------Ideal------------------------------------------ 1820 // If the load is from Field memory and the pointer is non-null, it might be possible to 1821 // zero out the control input. 1822 // If the offset is constant and the base is an object allocation, 1823 // try to hook me up to the exact initializing store. 1824 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1825 if (has_pinned_control_dependency()) { 1826 return nullptr; 1827 } 1828 Node* p = MemNode::Ideal_common(phase, can_reshape); 1829 if (p) return (p == NodeSentinel) ? nullptr : p; 1830 1831 Node* ctrl = in(MemNode::Control); 1832 Node* address = in(MemNode::Address); 1833 bool progress = false; 1834 1835 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) && 1836 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes()); 1837 1838 // Skip up past a SafePoint control. Cannot do this for Stores because 1839 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1840 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint && 1841 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw && 1842 !addr_mark && 1843 (depends_only_on_test() || has_unknown_control_dependency())) { 1844 ctrl = ctrl->in(0); 1845 set_req(MemNode::Control,ctrl); 1846 progress = true; 1847 } 1848 1849 intptr_t ignore = 0; 1850 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1851 if (base != nullptr 1852 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1853 // Check for useless control edge in some common special cases 1854 if (in(MemNode::Control) != nullptr 1855 && can_remove_control() 1856 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1857 && all_controls_dominate(base, phase->C->start())) { 1858 // A method-invariant, non-null address (constant or 'this' argument). 1859 set_req(MemNode::Control, nullptr); 1860 progress = true; 1861 } 1862 } 1863 1864 Node* mem = in(MemNode::Memory); 1865 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1866 1867 if (can_reshape && (addr_t != nullptr)) { 1868 // try to optimize our memory input 1869 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1870 if (opt_mem != mem) { 1871 set_req_X(MemNode::Memory, opt_mem, phase); 1872 if (phase->type( opt_mem ) == Type::TOP) return nullptr; 1873 return this; 1874 } 1875 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1876 if ((t_oop != nullptr) && 1877 (t_oop->is_known_instance_field() || 1878 t_oop->is_ptr_to_boxed_value())) { 1879 PhaseIterGVN *igvn = phase->is_IterGVN(); 1880 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true"); 1881 if (igvn->_worklist.member(opt_mem)) { 1882 // Delay this transformation until memory Phi is processed. 1883 igvn->_worklist.push(this); 1884 return nullptr; 1885 } 1886 // Split instance field load through Phi. 1887 Node* result = split_through_phi(phase); 1888 if (result != nullptr) return result; 1889 1890 if (t_oop->is_ptr_to_boxed_value()) { 1891 Node* result = eliminate_autobox(igvn); 1892 if (result != nullptr) return result; 1893 } 1894 } 1895 } 1896 1897 // Is there a dominating load that loads the same value? Leave 1898 // anything that is not a load of a field/array element (like 1899 // barriers etc.) alone 1900 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) { 1901 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1902 Node *use = mem->fast_out(i); 1903 if (use != this && 1904 use->Opcode() == Opcode() && 1905 use->in(0) != nullptr && 1906 use->in(0) != in(0) && 1907 use->in(Address) == in(Address)) { 1908 Node* ctl = in(0); 1909 for (int i = 0; i < 10 && ctl != nullptr; i++) { 1910 ctl = IfNode::up_one_dom(ctl); 1911 if (ctl == use->in(0)) { 1912 set_req(0, use->in(0)); 1913 return this; 1914 } 1915 } 1916 } 1917 } 1918 } 1919 1920 // Check for prior store with a different base or offset; make Load 1921 // independent. Skip through any number of them. Bail out if the stores 1922 // are in an endless dead cycle and report no progress. This is a key 1923 // transform for Reflection. However, if after skipping through the Stores 1924 // we can't then fold up against a prior store do NOT do the transform as 1925 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1926 // array memory alive twice: once for the hoisted Load and again after the 1927 // bypassed Store. This situation only works if EVERYBODY who does 1928 // anti-dependence work knows how to bypass. I.e. we need all 1929 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1930 // the alias index stuff. So instead, peek through Stores and IFF we can 1931 // fold up, do so. 1932 Node* prev_mem = find_previous_store(phase); 1933 if (prev_mem != nullptr) { 1934 Node* value = can_see_arraycopy_value(prev_mem, phase); 1935 if (value != nullptr) { 1936 return value; 1937 } 1938 } 1939 // Steps (a), (b): Walk past independent stores to find an exact match. 1940 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) { 1941 // (c) See if we can fold up on the spot, but don't fold up here. 1942 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1943 // just return a prior value, which is done by Identity calls. 1944 if (can_see_stored_value(prev_mem, phase)) { 1945 // Make ready for step (d): 1946 set_req_X(MemNode::Memory, prev_mem, phase); 1947 return this; 1948 } 1949 } 1950 1951 return progress ? this : nullptr; 1952 } 1953 1954 // Helper to recognize certain Klass fields which are invariant across 1955 // some group of array types (e.g., int[] or all T[] where T < Object). 1956 const Type* 1957 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1958 ciKlass* klass) const { 1959 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1960 // The field is Klass::_modifier_flags. Return its (constant) value. 1961 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1962 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1963 return TypeInt::make(klass->modifier_flags()); 1964 } 1965 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1966 // The field is Klass::_access_flags. Return its (constant) value. 1967 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1968 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1969 return TypeInt::make(klass->access_flags()); 1970 } 1971 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1972 // The field is Klass::_layout_helper. Return its constant value if known. 1973 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1974 return TypeInt::make(klass->layout_helper()); 1975 } 1976 1977 // No match. 1978 return nullptr; 1979 } 1980 1981 //------------------------------Value----------------------------------------- 1982 const Type* LoadNode::Value(PhaseGVN* phase) const { 1983 // Either input is TOP ==> the result is TOP 1984 Node* mem = in(MemNode::Memory); 1985 const Type *t1 = phase->type(mem); 1986 if (t1 == Type::TOP) return Type::TOP; 1987 Node* adr = in(MemNode::Address); 1988 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1989 if (tp == nullptr || tp->empty()) return Type::TOP; 1990 int off = tp->offset(); 1991 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1992 Compile* C = phase->C; 1993 1994 // Try to guess loaded type from pointer type 1995 if (tp->isa_aryptr()) { 1996 const TypeAryPtr* ary = tp->is_aryptr(); 1997 const Type* t = ary->elem(); 1998 1999 // Determine whether the reference is beyond the header or not, by comparing 2000 // the offset against the offset of the start of the array's data. 2001 // Different array types begin at slightly different offsets (12 vs. 16). 2002 // We choose T_BYTE as an example base type that is least restrictive 2003 // as to alignment, which will therefore produce the smallest 2004 // possible base offset. 2005 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 2006 const bool off_beyond_header = (off >= min_base_off); 2007 2008 // Try to constant-fold a stable array element. 2009 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) { 2010 // Make sure the reference is not into the header and the offset is constant 2011 ciObject* aobj = ary->const_oop(); 2012 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 2013 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0); 2014 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, 2015 stable_dimension, 2016 memory_type(), is_unsigned()); 2017 if (con_type != nullptr) { 2018 return con_type; 2019 } 2020 } 2021 } 2022 2023 // Don't do this for integer types. There is only potential profit if 2024 // the element type t is lower than _type; that is, for int types, if _type is 2025 // more restrictive than t. This only happens here if one is short and the other 2026 // char (both 16 bits), and in those cases we've made an intentional decision 2027 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 2028 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 2029 // 2030 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 2031 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 2032 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 2033 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 2034 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 2035 // In fact, that could have been the original type of p1, and p1 could have 2036 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 2037 // expression (LShiftL quux 3) independently optimized to the constant 8. 2038 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr) 2039 && (_type->isa_vect() == nullptr) 2040 && !ary->is_flat() 2041 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 2042 // t might actually be lower than _type, if _type is a unique 2043 // concrete subclass of abstract class t. 2044 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header? 2045 const Type* jt = t->join_speculative(_type); 2046 // In any case, do not allow the join, per se, to empty out the type. 2047 if (jt->empty() && !t->empty()) { 2048 // This can happen if a interface-typed array narrows to a class type. 2049 jt = _type; 2050 } 2051 #ifdef ASSERT 2052 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 2053 // The pointers in the autobox arrays are always non-null 2054 Node* base = adr->in(AddPNode::Base); 2055 if ((base != nullptr) && base->is_DecodeN()) { 2056 // Get LoadN node which loads IntegerCache.cache field 2057 base = base->in(1); 2058 } 2059 if ((base != nullptr) && base->is_Con()) { 2060 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 2061 if ((base_type != nullptr) && base_type->is_autobox_cache()) { 2062 // It could be narrow oop 2063 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 2064 } 2065 } 2066 } 2067 #endif 2068 return jt; 2069 } 2070 } 2071 } else if (tp->base() == Type::InstPtr) { 2072 assert( off != Type::OffsetBot || 2073 // arrays can be cast to Objects 2074 !tp->isa_instptr() || 2075 tp->is_instptr()->instance_klass()->is_java_lang_Object() || 2076 // Default value load 2077 tp->is_instptr()->instance_klass() == ciEnv::current()->Class_klass() || 2078 // unsafe field access may not have a constant offset 2079 C->has_unsafe_access(), 2080 "Field accesses must be precise" ); 2081 // For oop loads, we expect the _type to be precise. 2082 2083 const TypeInstPtr* tinst = tp->is_instptr(); 2084 BasicType bt = memory_type(); 2085 2086 // Optimize loads from constant fields. 2087 ciObject* const_oop = tinst->const_oop(); 2088 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) { 2089 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt); 2090 if (con_type != nullptr) { 2091 return con_type; 2092 } 2093 } 2094 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) { 2095 assert(off != Type::OffsetBot || 2096 !tp->isa_instklassptr() || 2097 // arrays can be cast to Objects 2098 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() || 2099 // also allow array-loading from the primary supertype 2100 // array during subtype checks 2101 Opcode() == Op_LoadKlass, 2102 "Field accesses must be precise"); 2103 // For klass/static loads, we expect the _type to be precise 2104 } else if (tp->base() == Type::RawPtr && !StressReflectiveCode) { 2105 if (adr->is_Load() && off == 0) { 2106 /* With mirrors being an indirect in the Klass* 2107 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset)) 2108 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass). 2109 * 2110 * So check the type and klass of the node before the LoadP. 2111 */ 2112 Node* adr2 = adr->in(MemNode::Address); 2113 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2114 if (tkls != nullptr) { 2115 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 2116 ciKlass* klass = tkls->exact_klass(); 2117 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 2118 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 2119 return TypeInstPtr::make(klass->java_mirror()); 2120 } 2121 } 2122 } else { 2123 // Check for a load of the default value offset from the InlineKlassFixedBlock: 2124 // LoadI(LoadP(inline_klass, adr_inlineklass_fixed_block_offset), default_value_offset_offset) 2125 // TODO 8325106 remove? 2126 intptr_t offset = 0; 2127 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2128 if (base != nullptr && base->is_Load() && offset == in_bytes(InlineKlass::default_value_offset_offset())) { 2129 const TypeKlassPtr* tkls = phase->type(base->in(MemNode::Address))->isa_klassptr(); 2130 if (tkls != nullptr && tkls->is_loaded() && tkls->klass_is_exact() && tkls->exact_klass()->is_inlinetype() && 2131 tkls->offset() == in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset())) { 2132 assert(base->Opcode() == Op_LoadP, "must load an oop from klass"); 2133 assert(Opcode() == Op_LoadI, "must load an int from fixed block"); 2134 return TypeInt::make(tkls->exact_klass()->as_inline_klass()->default_value_offset()); 2135 } 2136 } 2137 } 2138 } 2139 2140 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2141 if (tkls != nullptr) { 2142 if (tkls->is_loaded() && tkls->klass_is_exact()) { 2143 ciKlass* klass = tkls->exact_klass(); 2144 // We are loading a field from a Klass metaobject whose identity 2145 // is known at compile time (the type is "exact" or "precise"). 2146 // Check for fields we know are maintained as constants by the VM. 2147 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 2148 // The field is Klass::_super_check_offset. Return its (constant) value. 2149 // (Folds up type checking code.) 2150 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 2151 return TypeInt::make(klass->super_check_offset()); 2152 } 2153 // Compute index into primary_supers array 2154 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 2155 // Check for overflowing; use unsigned compare to handle the negative case. 2156 if( depth < ciKlass::primary_super_limit() ) { 2157 // The field is an element of Klass::_primary_supers. Return its (constant) value. 2158 // (Folds up type checking code.) 2159 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 2160 ciKlass *ss = klass->super_of_depth(depth); 2161 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR; 2162 } 2163 const Type* aift = load_array_final_field(tkls, klass); 2164 if (aift != nullptr) return aift; 2165 } 2166 2167 // We can still check if we are loading from the primary_supers array at a 2168 // shallow enough depth. Even though the klass is not exact, entries less 2169 // than or equal to its super depth are correct. 2170 if (tkls->is_loaded()) { 2171 ciKlass* klass = nullptr; 2172 if (tkls->isa_instklassptr()) { 2173 klass = tkls->is_instklassptr()->instance_klass(); 2174 } else { 2175 int dims; 2176 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims); 2177 if (inner->isa_instklassptr()) { 2178 klass = inner->is_instklassptr()->instance_klass(); 2179 klass = ciObjArrayKlass::make(klass, dims); 2180 } 2181 } 2182 if (klass != nullptr) { 2183 // Compute index into primary_supers array 2184 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 2185 // Check for overflowing; use unsigned compare to handle the negative case. 2186 if (depth < ciKlass::primary_super_limit() && 2187 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case 2188 // The field is an element of Klass::_primary_supers. Return its (constant) value. 2189 // (Folds up type checking code.) 2190 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 2191 ciKlass *ss = klass->super_of_depth(depth); 2192 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR; 2193 } 2194 } 2195 } 2196 2197 // If the type is enough to determine that the thing is not an array, 2198 // we can give the layout_helper a positive interval type. 2199 // This will help short-circuit some reflective code. 2200 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) && 2201 tkls->isa_instklassptr() && // not directly typed as an array 2202 !tkls->is_instklassptr()->instance_klass()->is_java_lang_Object() // not the supertype of all T[] and specifically not Serializable & Cloneable 2203 ) { 2204 // Note: When interfaces are reliable, we can narrow the interface 2205 // test to (klass != Serializable && klass != Cloneable). 2206 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 2207 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 2208 // The key property of this type is that it folds up tests 2209 // for array-ness, since it proves that the layout_helper is positive. 2210 // Thus, a generic value like the basic object layout helper works fine. 2211 return TypeInt::make(min_size, max_jint, Type::WidenMin); 2212 } 2213 } 2214 2215 // If we are loading from a freshly-allocated object, produce a zero, 2216 // if the load is provably beyond the header of the object. 2217 // (Also allow a variable load from a fresh array to produce zero.) 2218 const TypeOopPtr *tinst = tp->isa_oopptr(); 2219 bool is_instance = (tinst != nullptr) && tinst->is_known_instance_field(); 2220 bool is_boxed_value = (tinst != nullptr) && tinst->is_ptr_to_boxed_value(); 2221 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 2222 Node* value = can_see_stored_value(mem,phase); 2223 if (value != nullptr && value->is_Con()) { 2224 assert(value->bottom_type()->higher_equal(_type),"sanity"); 2225 return value->bottom_type(); 2226 } 2227 } 2228 2229 bool is_vect = (_type->isa_vect() != nullptr); 2230 if (is_instance && !is_vect) { 2231 // If we have an instance type and our memory input is the 2232 // programs's initial memory state, there is no matching store, 2233 // so just return a zero of the appropriate type - 2234 // except if it is vectorized - then we have no zero constant. 2235 Node *mem = in(MemNode::Memory); 2236 if (mem->is_Parm() && mem->in(0)->is_Start()) { 2237 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 2238 return Type::get_zero_type(_type->basic_type()); 2239 } 2240 } 2241 Node* alloc = is_new_object_mark_load(); 2242 if (alloc != nullptr) { 2243 if (EnableValhalla) { 2244 // The mark word may contain property bits (inline, flat, null-free) 2245 Node* klass_node = alloc->in(AllocateNode::KlassNode); 2246 const TypeKlassPtr* tkls = phase->type(klass_node)->isa_klassptr(); 2247 if (tkls != nullptr && tkls->is_loaded() && tkls->klass_is_exact()) { 2248 return TypeX::make(tkls->exact_klass()->prototype_header().value()); 2249 } 2250 } else { 2251 return TypeX::make(markWord::prototype().value()); 2252 } 2253 } 2254 2255 return _type; 2256 } 2257 2258 //------------------------------match_edge------------------------------------- 2259 // Do we Match on this edge index or not? Match only the address. 2260 uint LoadNode::match_edge(uint idx) const { 2261 return idx == MemNode::Address; 2262 } 2263 2264 //--------------------------LoadBNode::Ideal-------------------------------------- 2265 // 2266 // If the previous store is to the same address as this load, 2267 // and the value stored was larger than a byte, replace this load 2268 // with the value stored truncated to a byte. If no truncation is 2269 // needed, the replacement is done in LoadNode::Identity(). 2270 // 2271 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2272 Node* mem = in(MemNode::Memory); 2273 Node* value = can_see_stored_value(mem,phase); 2274 if (value != nullptr) { 2275 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false); 2276 if (narrow != value) { 2277 return narrow; 2278 } 2279 } 2280 // Identity call will handle the case where truncation is not needed. 2281 return LoadNode::Ideal(phase, can_reshape); 2282 } 2283 2284 const Type* LoadBNode::Value(PhaseGVN* phase) const { 2285 Node* mem = in(MemNode::Memory); 2286 Node* value = can_see_stored_value(mem,phase); 2287 if (value != nullptr && value->is_Con() && 2288 !value->bottom_type()->higher_equal(_type)) { 2289 // If the input to the store does not fit with the load's result type, 2290 // it must be truncated. We can't delay until Ideal call since 2291 // a singleton Value is needed for split_thru_phi optimization. 2292 int con = value->get_int(); 2293 return TypeInt::make((con << 24) >> 24); 2294 } 2295 return LoadNode::Value(phase); 2296 } 2297 2298 //--------------------------LoadUBNode::Ideal------------------------------------- 2299 // 2300 // If the previous store is to the same address as this load, 2301 // and the value stored was larger than a byte, replace this load 2302 // with the value stored truncated to a byte. If no truncation is 2303 // needed, the replacement is done in LoadNode::Identity(). 2304 // 2305 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2306 Node* mem = in(MemNode::Memory); 2307 Node* value = can_see_stored_value(mem, phase); 2308 if (value != nullptr) { 2309 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false); 2310 if (narrow != value) { 2311 return narrow; 2312 } 2313 } 2314 // Identity call will handle the case where truncation is not needed. 2315 return LoadNode::Ideal(phase, can_reshape); 2316 } 2317 2318 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 2319 Node* mem = in(MemNode::Memory); 2320 Node* value = can_see_stored_value(mem,phase); 2321 if (value != nullptr && value->is_Con() && 2322 !value->bottom_type()->higher_equal(_type)) { 2323 // If the input to the store does not fit with the load's result type, 2324 // it must be truncated. We can't delay until Ideal call since 2325 // a singleton Value is needed for split_thru_phi optimization. 2326 int con = value->get_int(); 2327 return TypeInt::make(con & 0xFF); 2328 } 2329 return LoadNode::Value(phase); 2330 } 2331 2332 //--------------------------LoadUSNode::Ideal------------------------------------- 2333 // 2334 // If the previous store is to the same address as this load, 2335 // and the value stored was larger than a char, replace this load 2336 // with the value stored truncated to a char. If no truncation is 2337 // needed, the replacement is done in LoadNode::Identity(). 2338 // 2339 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2340 Node* mem = in(MemNode::Memory); 2341 Node* value = can_see_stored_value(mem,phase); 2342 if (value != nullptr) { 2343 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false); 2344 if (narrow != value) { 2345 return narrow; 2346 } 2347 } 2348 // Identity call will handle the case where truncation is not needed. 2349 return LoadNode::Ideal(phase, can_reshape); 2350 } 2351 2352 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 2353 Node* mem = in(MemNode::Memory); 2354 Node* value = can_see_stored_value(mem,phase); 2355 if (value != nullptr && value->is_Con() && 2356 !value->bottom_type()->higher_equal(_type)) { 2357 // If the input to the store does not fit with the load's result type, 2358 // it must be truncated. We can't delay until Ideal call since 2359 // a singleton Value is needed for split_thru_phi optimization. 2360 int con = value->get_int(); 2361 return TypeInt::make(con & 0xFFFF); 2362 } 2363 return LoadNode::Value(phase); 2364 } 2365 2366 //--------------------------LoadSNode::Ideal-------------------------------------- 2367 // 2368 // If the previous store is to the same address as this load, 2369 // and the value stored was larger than a short, replace this load 2370 // with the value stored truncated to a short. If no truncation is 2371 // needed, the replacement is done in LoadNode::Identity(). 2372 // 2373 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2374 Node* mem = in(MemNode::Memory); 2375 Node* value = can_see_stored_value(mem,phase); 2376 if (value != nullptr) { 2377 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false); 2378 if (narrow != value) { 2379 return narrow; 2380 } 2381 } 2382 // Identity call will handle the case where truncation is not needed. 2383 return LoadNode::Ideal(phase, can_reshape); 2384 } 2385 2386 const Type* LoadSNode::Value(PhaseGVN* phase) const { 2387 Node* mem = in(MemNode::Memory); 2388 Node* value = can_see_stored_value(mem,phase); 2389 if (value != nullptr && value->is_Con() && 2390 !value->bottom_type()->higher_equal(_type)) { 2391 // If the input to the store does not fit with the load's result type, 2392 // it must be truncated. We can't delay until Ideal call since 2393 // a singleton Value is needed for split_thru_phi optimization. 2394 int con = value->get_int(); 2395 return TypeInt::make((con << 16) >> 16); 2396 } 2397 return LoadNode::Value(phase); 2398 } 2399 2400 //============================================================================= 2401 //----------------------------LoadKlassNode::make------------------------------ 2402 // Polymorphic factory method: 2403 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, 2404 const TypeKlassPtr* tk) { 2405 // sanity check the alias category against the created node type 2406 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2407 assert(adr_type != nullptr, "expecting TypeKlassPtr"); 2408 #ifdef _LP64 2409 if (adr_type->is_ptr_to_narrowklass()) { 2410 assert(UseCompressedClassPointers, "no compressed klasses"); 2411 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2412 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2413 } 2414 #endif 2415 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2416 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 2417 } 2418 2419 //------------------------------Value------------------------------------------ 2420 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 2421 return klass_value_common(phase); 2422 } 2423 2424 // In most cases, LoadKlassNode does not have the control input set. If the control 2425 // input is set, it must not be removed (by LoadNode::Ideal()). 2426 bool LoadKlassNode::can_remove_control() const { 2427 return false; 2428 } 2429 2430 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2431 // Either input is TOP ==> the result is TOP 2432 const Type *t1 = phase->type( in(MemNode::Memory) ); 2433 if (t1 == Type::TOP) return Type::TOP; 2434 Node *adr = in(MemNode::Address); 2435 const Type *t2 = phase->type( adr ); 2436 if (t2 == Type::TOP) return Type::TOP; 2437 const TypePtr *tp = t2->is_ptr(); 2438 if (TypePtr::above_centerline(tp->ptr()) || 2439 tp->ptr() == TypePtr::Null) return Type::TOP; 2440 2441 // Return a more precise klass, if possible 2442 const TypeInstPtr *tinst = tp->isa_instptr(); 2443 if (tinst != nullptr) { 2444 ciInstanceKlass* ik = tinst->instance_klass(); 2445 int offset = tinst->offset(); 2446 if (ik == phase->C->env()->Class_klass() 2447 && (offset == java_lang_Class::klass_offset() || 2448 offset == java_lang_Class::array_klass_offset())) { 2449 // We are loading a special hidden field from a Class mirror object, 2450 // the field which points to the VM's Klass metaobject. 2451 bool is_null_free_array = false; 2452 ciType* t = tinst->java_mirror_type(&is_null_free_array); 2453 // java_mirror_type returns non-null for compile-time Class constants. 2454 if (t != nullptr) { 2455 // constant oop => constant klass 2456 if (offset == java_lang_Class::array_klass_offset()) { 2457 if (t->is_void()) { 2458 // We cannot create a void array. Since void is a primitive type return null 2459 // klass. Users of this result need to do a null check on the returned klass. 2460 return TypePtr::NULL_PTR; 2461 } 2462 const TypeKlassPtr* tklass = TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces); 2463 if (is_null_free_array) { 2464 tklass = tklass->is_aryklassptr()->cast_to_null_free(); 2465 } 2466 return tklass; 2467 } 2468 if (!t->is_klass()) { 2469 // a primitive Class (e.g., int.class) has null for a klass field 2470 return TypePtr::NULL_PTR; 2471 } 2472 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2473 const TypeKlassPtr* tklass = TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces); 2474 if (is_null_free_array) { 2475 tklass = tklass->is_aryklassptr()->cast_to_null_free(); 2476 } 2477 return tklass; 2478 } 2479 // non-constant mirror, so we can't tell what's going on 2480 } 2481 if (!tinst->is_loaded()) 2482 return _type; // Bail out if not loaded 2483 if (offset == oopDesc::klass_offset_in_bytes()) { 2484 return tinst->as_klass_type(true); 2485 } 2486 } 2487 2488 // Check for loading klass from an array 2489 const TypeAryPtr* tary = tp->isa_aryptr(); 2490 if (tary != nullptr && 2491 tary->offset() == oopDesc::klass_offset_in_bytes()) { 2492 return tary->as_klass_type(true); 2493 } 2494 2495 // Check for loading klass from an array klass 2496 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2497 if (tkls != nullptr && !StressReflectiveCode) { 2498 if (!tkls->is_loaded()) 2499 return _type; // Bail out if not loaded 2500 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() && 2501 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2502 // // Always returning precise element type is incorrect, 2503 // // e.g., element type could be object and array may contain strings 2504 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2505 2506 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2507 // according to the element type's subclassing. 2508 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact()); 2509 } 2510 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() && 2511 tkls->offset() == in_bytes(Klass::super_offset())) { 2512 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super(); 2513 // The field is Klass::_super. Return its (constant) value. 2514 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2515 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR; 2516 } 2517 } 2518 2519 // Bailout case 2520 return LoadNode::Value(phase); 2521 } 2522 2523 //------------------------------Identity--------------------------------------- 2524 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2525 // Also feed through the klass in Allocate(...klass...)._klass. 2526 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2527 return klass_identity_common(phase); 2528 } 2529 2530 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2531 Node* x = LoadNode::Identity(phase); 2532 if (x != this) return x; 2533 2534 // Take apart the address into an oop and offset. 2535 // Return 'this' if we cannot. 2536 Node* adr = in(MemNode::Address); 2537 intptr_t offset = 0; 2538 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2539 if (base == nullptr) return this; 2540 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2541 if (toop == nullptr) return this; 2542 2543 // Step over potential GC barrier for OopHandle resolve 2544 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2545 if (bs->is_gc_barrier_node(base)) { 2546 base = bs->step_over_gc_barrier(base); 2547 } 2548 2549 // We can fetch the klass directly through an AllocateNode. 2550 // This works even if the klass is not constant (clone or newArray). 2551 if (offset == oopDesc::klass_offset_in_bytes()) { 2552 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2553 if (allocated_klass != nullptr) { 2554 return allocated_klass; 2555 } 2556 } 2557 2558 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2559 // See inline_native_Class_query for occurrences of these patterns. 2560 // Java Example: x.getClass().isAssignableFrom(y) 2561 // 2562 // This improves reflective code, often making the Class 2563 // mirror go completely dead. (Current exception: Class 2564 // mirrors may appear in debug info, but we could clean them out by 2565 // introducing a new debug info operator for Klass.java_mirror). 2566 2567 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass() 2568 && offset == java_lang_Class::klass_offset()) { 2569 if (base->is_Load()) { 2570 Node* base2 = base->in(MemNode::Address); 2571 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */ 2572 Node* adr2 = base2->in(MemNode::Address); 2573 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2574 if (tkls != nullptr && !tkls->empty() 2575 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr()) 2576 && adr2->is_AddP() 2577 ) { 2578 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2579 if (tkls->offset() == mirror_field) { 2580 return adr2->in(AddPNode::Base); 2581 } 2582 } 2583 } 2584 } 2585 } 2586 2587 return this; 2588 } 2589 2590 LoadNode* LoadNode::clone_pinned() const { 2591 LoadNode* ld = clone()->as_Load(); 2592 ld->_control_dependency = UnknownControl; 2593 return ld; 2594 } 2595 2596 2597 //------------------------------Value------------------------------------------ 2598 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2599 const Type *t = klass_value_common(phase); 2600 if (t == Type::TOP) 2601 return t; 2602 2603 return t->make_narrowklass(); 2604 } 2605 2606 //------------------------------Identity--------------------------------------- 2607 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2608 // Also feed through the klass in Allocate(...klass...)._klass. 2609 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2610 Node *x = klass_identity_common(phase); 2611 2612 const Type *t = phase->type( x ); 2613 if( t == Type::TOP ) return x; 2614 if( t->isa_narrowklass()) return x; 2615 assert (!t->isa_narrowoop(), "no narrow oop here"); 2616 2617 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2618 } 2619 2620 //------------------------------Value----------------------------------------- 2621 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2622 // Either input is TOP ==> the result is TOP 2623 const Type *t1 = phase->type( in(MemNode::Memory) ); 2624 if( t1 == Type::TOP ) return Type::TOP; 2625 Node *adr = in(MemNode::Address); 2626 const Type *t2 = phase->type( adr ); 2627 if( t2 == Type::TOP ) return Type::TOP; 2628 const TypePtr *tp = t2->is_ptr(); 2629 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2630 const TypeAryPtr *tap = tp->isa_aryptr(); 2631 if( !tap ) return _type; 2632 return tap->size(); 2633 } 2634 2635 //-------------------------------Ideal--------------------------------------- 2636 // Feed through the length in AllocateArray(...length...)._length. 2637 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2638 Node* p = MemNode::Ideal_common(phase, can_reshape); 2639 if (p) return (p == NodeSentinel) ? nullptr : p; 2640 2641 // Take apart the address into an oop and offset. 2642 // Return 'this' if we cannot. 2643 Node* adr = in(MemNode::Address); 2644 intptr_t offset = 0; 2645 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2646 if (base == nullptr) return nullptr; 2647 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2648 if (tary == nullptr) return nullptr; 2649 2650 // We can fetch the length directly through an AllocateArrayNode. 2651 // This works even if the length is not constant (clone or newArray). 2652 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2653 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base); 2654 if (alloc != nullptr) { 2655 Node* allocated_length = alloc->Ideal_length(); 2656 Node* len = alloc->make_ideal_length(tary, phase); 2657 if (allocated_length != len) { 2658 // New CastII improves on this. 2659 return len; 2660 } 2661 } 2662 } 2663 2664 return nullptr; 2665 } 2666 2667 //------------------------------Identity--------------------------------------- 2668 // Feed through the length in AllocateArray(...length...)._length. 2669 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2670 Node* x = LoadINode::Identity(phase); 2671 if (x != this) return x; 2672 2673 // Take apart the address into an oop and offset. 2674 // Return 'this' if we cannot. 2675 Node* adr = in(MemNode::Address); 2676 intptr_t offset = 0; 2677 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2678 if (base == nullptr) return this; 2679 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2680 if (tary == nullptr) return this; 2681 2682 // We can fetch the length directly through an AllocateArrayNode. 2683 // This works even if the length is not constant (clone or newArray). 2684 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2685 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base); 2686 if (alloc != nullptr) { 2687 Node* allocated_length = alloc->Ideal_length(); 2688 // Do not allow make_ideal_length to allocate a CastII node. 2689 Node* len = alloc->make_ideal_length(tary, phase, false); 2690 if (allocated_length == len) { 2691 // Return allocated_length only if it would not be improved by a CastII. 2692 return allocated_length; 2693 } 2694 } 2695 } 2696 2697 return this; 2698 2699 } 2700 2701 //============================================================================= 2702 //---------------------------StoreNode::make----------------------------------- 2703 // Polymorphic factory method: 2704 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) { 2705 assert((mo == unordered || mo == release), "unexpected"); 2706 Compile* C = gvn.C; 2707 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2708 ctl != nullptr, "raw memory operations should have control edge"); 2709 2710 switch (bt) { 2711 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2712 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2713 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2714 case T_CHAR: 2715 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2716 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access); 2717 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2718 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access); 2719 case T_METADATA: 2720 case T_ADDRESS: 2721 case T_OBJECT: 2722 #ifdef _LP64 2723 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2724 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2725 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2726 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2727 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2728 adr->bottom_type()->isa_rawptr())) { 2729 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2730 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2731 } 2732 #endif 2733 { 2734 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2735 } 2736 default: 2737 ShouldNotReachHere(); 2738 return (StoreNode*)nullptr; 2739 } 2740 } 2741 2742 //--------------------------bottom_type---------------------------------------- 2743 const Type *StoreNode::bottom_type() const { 2744 return Type::MEMORY; 2745 } 2746 2747 //------------------------------hash------------------------------------------- 2748 uint StoreNode::hash() const { 2749 // unroll addition of interesting fields 2750 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2751 2752 // Since they are not commoned, do not hash them: 2753 return NO_HASH; 2754 } 2755 2756 //------------------------------Ideal------------------------------------------ 2757 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2758 // When a store immediately follows a relevant allocation/initialization, 2759 // try to capture it into the initialization, or hoist it above. 2760 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2761 Node* p = MemNode::Ideal_common(phase, can_reshape); 2762 if (p) return (p == NodeSentinel) ? nullptr : p; 2763 2764 Node* mem = in(MemNode::Memory); 2765 Node* address = in(MemNode::Address); 2766 Node* value = in(MemNode::ValueIn); 2767 // Back-to-back stores to same address? Fold em up. Generally 2768 // unsafe if I have intervening uses... Also disallowed for StoreCM 2769 // since they must follow each StoreP operation. Redundant StoreCMs 2770 // are eliminated just before matching in final_graph_reshape. 2771 if (phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::INLINES) { 2772 Node* st = mem; 2773 // If Store 'st' has more than one use, we cannot fold 'st' away. 2774 // For example, 'st' might be the final state at a conditional 2775 // return. Or, 'st' might be used by some node which is live at 2776 // the same time 'st' is live, which might be unschedulable. So, 2777 // require exactly ONE user until such time as we clone 'mem' for 2778 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2779 // true). 2780 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2781 // Looking at a dead closed cycle of memory? 2782 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2783 assert(Opcode() == st->Opcode() || 2784 st->Opcode() == Op_StoreVector || 2785 Opcode() == Op_StoreVector || 2786 st->Opcode() == Op_StoreVectorScatter || 2787 Opcode() == Op_StoreVectorScatter || 2788 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2789 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 2790 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy 2791 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) || 2792 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 2793 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2794 2795 if (st->in(MemNode::Address)->eqv_uncast(address) && 2796 st->as_Store()->memory_size() <= this->memory_size()) { 2797 Node* use = st->raw_out(0); 2798 if (phase->is_IterGVN()) { 2799 phase->is_IterGVN()->rehash_node_delayed(use); 2800 } 2801 // It's OK to do this in the parser, since DU info is always accurate, 2802 // and the parser always refers to nodes via SafePointNode maps. 2803 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase); 2804 return this; 2805 } 2806 st = st->in(MemNode::Memory); 2807 } 2808 } 2809 2810 2811 // Capture an unaliased, unconditional, simple store into an initializer. 2812 // Or, if it is independent of the allocation, hoist it above the allocation. 2813 if (ReduceFieldZeroing && /*can_reshape &&*/ 2814 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2815 InitializeNode* init = mem->in(0)->as_Initialize(); 2816 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2817 if (offset > 0) { 2818 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2819 // If the InitializeNode captured me, it made a raw copy of me, 2820 // and I need to disappear. 2821 if (moved != nullptr) { 2822 // %%% hack to ensure that Ideal returns a new node: 2823 mem = MergeMemNode::make(mem); 2824 return mem; // fold me away 2825 } 2826 } 2827 } 2828 2829 // Fold reinterpret cast into memory operation: 2830 // StoreX mem (MoveY2X v) => StoreY mem v 2831 if (value->is_Move()) { 2832 const Type* vt = value->in(1)->bottom_type(); 2833 if (has_reinterpret_variant(vt)) { 2834 if (phase->C->post_loop_opts_phase()) { 2835 return convert_to_reinterpret_store(*phase, value->in(1), vt); 2836 } else { 2837 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over 2838 } 2839 } 2840 } 2841 2842 return nullptr; // No further progress 2843 } 2844 2845 //------------------------------Value----------------------------------------- 2846 const Type* StoreNode::Value(PhaseGVN* phase) const { 2847 // Either input is TOP ==> the result is TOP 2848 const Type *t1 = phase->type( in(MemNode::Memory) ); 2849 if( t1 == Type::TOP ) return Type::TOP; 2850 const Type *t2 = phase->type( in(MemNode::Address) ); 2851 if( t2 == Type::TOP ) return Type::TOP; 2852 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2853 if( t3 == Type::TOP ) return Type::TOP; 2854 return Type::MEMORY; 2855 } 2856 2857 //------------------------------Identity--------------------------------------- 2858 // Remove redundant stores: 2859 // Store(m, p, Load(m, p)) changes to m. 2860 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2861 Node* StoreNode::Identity(PhaseGVN* phase) { 2862 Node* mem = in(MemNode::Memory); 2863 Node* adr = in(MemNode::Address); 2864 Node* val = in(MemNode::ValueIn); 2865 2866 Node* result = this; 2867 2868 // Load then Store? Then the Store is useless 2869 if (val->is_Load() && 2870 val->in(MemNode::Address)->eqv_uncast(adr) && 2871 val->in(MemNode::Memory )->eqv_uncast(mem) && 2872 val->as_Load()->store_Opcode() == Opcode()) { 2873 result = mem; 2874 } 2875 2876 // Two stores in a row of the same value? 2877 if (result == this && 2878 mem->is_Store() && 2879 mem->in(MemNode::Address)->eqv_uncast(adr) && 2880 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2881 mem->Opcode() == Opcode()) { 2882 result = mem; 2883 } 2884 2885 // Store of zero anywhere into a freshly-allocated object? 2886 // Then the store is useless. 2887 // (It must already have been captured by the InitializeNode.) 2888 if (result == this && ReduceFieldZeroing) { 2889 // a newly allocated object is already all-zeroes everywhere 2890 if (mem->is_Proj() && mem->in(0)->is_Allocate() && 2891 (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == val)) { 2892 result = mem; 2893 } 2894 2895 if (result == this && phase->type(val)->is_zero_type()) { 2896 // the store may also apply to zero-bits in an earlier object 2897 Node* prev_mem = find_previous_store(phase); 2898 // Steps (a), (b): Walk past independent stores to find an exact match. 2899 if (prev_mem != nullptr) { 2900 Node* prev_val = can_see_stored_value(prev_mem, phase); 2901 if (prev_val != nullptr && prev_val == val) { 2902 // prev_val and val might differ by a cast; it would be good 2903 // to keep the more informative of the two. 2904 result = mem; 2905 } 2906 } 2907 } 2908 } 2909 2910 PhaseIterGVN* igvn = phase->is_IterGVN(); 2911 if (result != this && igvn != nullptr) { 2912 MemBarNode* trailing = trailing_membar(); 2913 if (trailing != nullptr) { 2914 #ifdef ASSERT 2915 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr(); 2916 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects"); 2917 #endif 2918 trailing->remove(igvn); 2919 } 2920 } 2921 2922 return result; 2923 } 2924 2925 //------------------------------match_edge------------------------------------- 2926 // Do we Match on this edge index or not? Match only memory & value 2927 uint StoreNode::match_edge(uint idx) const { 2928 return idx == MemNode::Address || idx == MemNode::ValueIn; 2929 } 2930 2931 //------------------------------cmp-------------------------------------------- 2932 // Do not common stores up together. They generally have to be split 2933 // back up anyways, so do not bother. 2934 bool StoreNode::cmp( const Node &n ) const { 2935 return (&n == this); // Always fail except on self 2936 } 2937 2938 //------------------------------Ideal_masked_input----------------------------- 2939 // Check for a useless mask before a partial-word store 2940 // (StoreB ... (AndI valIn conIa) ) 2941 // If (conIa & mask == mask) this simplifies to 2942 // (StoreB ... (valIn) ) 2943 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2944 Node *val = in(MemNode::ValueIn); 2945 if( val->Opcode() == Op_AndI ) { 2946 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2947 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2948 set_req_X(MemNode::ValueIn, val->in(1), phase); 2949 return this; 2950 } 2951 } 2952 return nullptr; 2953 } 2954 2955 2956 //------------------------------Ideal_sign_extended_input---------------------- 2957 // Check for useless sign-extension before a partial-word store 2958 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2959 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2960 // (StoreB ... (valIn) ) 2961 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2962 Node *val = in(MemNode::ValueIn); 2963 if( val->Opcode() == Op_RShiftI ) { 2964 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2965 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2966 Node *shl = val->in(1); 2967 if( shl->Opcode() == Op_LShiftI ) { 2968 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2969 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2970 set_req_X(MemNode::ValueIn, shl->in(1), phase); 2971 return this; 2972 } 2973 } 2974 } 2975 } 2976 return nullptr; 2977 } 2978 2979 //------------------------------value_never_loaded----------------------------------- 2980 // Determine whether there are any possible loads of the value stored. 2981 // For simplicity, we actually check if there are any loads from the 2982 // address stored to, not just for loads of the value stored by this node. 2983 // 2984 bool StoreNode::value_never_loaded(PhaseValues* phase) const { 2985 Node *adr = in(Address); 2986 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2987 if (adr_oop == nullptr) 2988 return false; 2989 if (!adr_oop->is_known_instance_field()) 2990 return false; // if not a distinct instance, there may be aliases of the address 2991 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2992 Node *use = adr->fast_out(i); 2993 if (use->is_Load() || use->is_LoadStore()) { 2994 return false; 2995 } 2996 } 2997 return true; 2998 } 2999 3000 MemBarNode* StoreNode::trailing_membar() const { 3001 if (is_release()) { 3002 MemBarNode* trailing_mb = nullptr; 3003 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 3004 Node* u = fast_out(i); 3005 if (u->is_MemBar()) { 3006 if (u->as_MemBar()->trailing_store()) { 3007 assert(u->Opcode() == Op_MemBarVolatile, ""); 3008 assert(trailing_mb == nullptr, "only one"); 3009 trailing_mb = u->as_MemBar(); 3010 #ifdef ASSERT 3011 Node* leading = u->as_MemBar()->leading_membar(); 3012 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 3013 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair"); 3014 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair"); 3015 #endif 3016 } else { 3017 assert(u->as_MemBar()->standalone(), ""); 3018 } 3019 } 3020 } 3021 return trailing_mb; 3022 } 3023 return nullptr; 3024 } 3025 3026 3027 //============================================================================= 3028 //------------------------------Ideal------------------------------------------ 3029 // If the store is from an AND mask that leaves the low bits untouched, then 3030 // we can skip the AND operation. If the store is from a sign-extension 3031 // (a left shift, then right shift) we can skip both. 3032 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 3033 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 3034 if( progress != nullptr ) return progress; 3035 3036 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 3037 if( progress != nullptr ) return progress; 3038 3039 // Finally check the default case 3040 return StoreNode::Ideal(phase, can_reshape); 3041 } 3042 3043 //============================================================================= 3044 //------------------------------Ideal------------------------------------------ 3045 // If the store is from an AND mask that leaves the low bits untouched, then 3046 // we can skip the AND operation 3047 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 3048 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 3049 if( progress != nullptr ) return progress; 3050 3051 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 3052 if( progress != nullptr ) return progress; 3053 3054 // Finally check the default case 3055 return StoreNode::Ideal(phase, can_reshape); 3056 } 3057 3058 //============================================================================= 3059 //------------------------------Identity--------------------------------------- 3060 Node* StoreCMNode::Identity(PhaseGVN* phase) { 3061 // No need to card mark when storing a null ptr 3062 Node* my_store = in(MemNode::OopStore); 3063 if (my_store->is_Store()) { 3064 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 3065 if( t1 == TypePtr::NULL_PTR ) { 3066 return in(MemNode::Memory); 3067 } 3068 } 3069 return this; 3070 } 3071 3072 //============================================================================= 3073 //------------------------------Ideal--------------------------------------- 3074 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 3075 Node* progress = StoreNode::Ideal(phase, can_reshape); 3076 if (progress != nullptr) return progress; 3077 3078 Node* my_store = in(MemNode::OopStore); 3079 if (my_store->is_MergeMem()) { 3080 if (oop_alias_idx() != phase->C->get_alias_index(TypeAryPtr::INLINES) || 3081 phase->C->flat_accesses_share_alias()) { 3082 // The alias that was recorded is no longer accurate enough. 3083 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 3084 set_req_X(MemNode::OopStore, mem, phase); 3085 return this; 3086 } 3087 } 3088 3089 return nullptr; 3090 } 3091 3092 //------------------------------Value----------------------------------------- 3093 const Type* StoreCMNode::Value(PhaseGVN* phase) const { 3094 // Either input is TOP ==> the result is TOP (checked in StoreNode::Value). 3095 // If extra input is TOP ==> the result is TOP 3096 const Type* t = phase->type(in(MemNode::OopStore)); 3097 if (t == Type::TOP) { 3098 return Type::TOP; 3099 } 3100 return StoreNode::Value(phase); 3101 } 3102 3103 3104 //============================================================================= 3105 //----------------------------------SCMemProjNode------------------------------ 3106 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 3107 { 3108 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) { 3109 return Type::TOP; 3110 } 3111 return bottom_type(); 3112 } 3113 3114 //============================================================================= 3115 //----------------------------------LoadStoreNode------------------------------ 3116 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 3117 : Node(required), 3118 _type(rt), 3119 _adr_type(at), 3120 _barrier_data(0) 3121 { 3122 init_req(MemNode::Control, c ); 3123 init_req(MemNode::Memory , mem); 3124 init_req(MemNode::Address, adr); 3125 init_req(MemNode::ValueIn, val); 3126 init_class_id(Class_LoadStore); 3127 } 3128 3129 //------------------------------Value----------------------------------------- 3130 const Type* LoadStoreNode::Value(PhaseGVN* phase) const { 3131 // Either input is TOP ==> the result is TOP 3132 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) { 3133 return Type::TOP; 3134 } 3135 const Type* t = phase->type(in(MemNode::Memory)); 3136 if (t == Type::TOP) { 3137 return Type::TOP; 3138 } 3139 t = phase->type(in(MemNode::Address)); 3140 if (t == Type::TOP) { 3141 return Type::TOP; 3142 } 3143 t = phase->type(in(MemNode::ValueIn)); 3144 if (t == Type::TOP) { 3145 return Type::TOP; 3146 } 3147 return bottom_type(); 3148 } 3149 3150 uint LoadStoreNode::ideal_reg() const { 3151 return _type->ideal_reg(); 3152 } 3153 3154 // This method conservatively checks if the result of a LoadStoreNode is 3155 // used, that is, if it returns true, then it is definitely the case that 3156 // the result of the node is not needed. 3157 // For example, GetAndAdd can be matched into a lock_add instead of a 3158 // lock_xadd if the result of LoadStoreNode::result_not_used() is true 3159 bool LoadStoreNode::result_not_used() const { 3160 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 3161 Node *x = fast_out(i); 3162 if (x->Opcode() == Op_SCMemProj) { 3163 continue; 3164 } 3165 if (x->bottom_type() == TypeTuple::MEMBAR && 3166 !x->is_Call() && 3167 x->Opcode() != Op_Blackhole) { 3168 continue; 3169 } 3170 return false; 3171 } 3172 return true; 3173 } 3174 3175 MemBarNode* LoadStoreNode::trailing_membar() const { 3176 MemBarNode* trailing = nullptr; 3177 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 3178 Node* u = fast_out(i); 3179 if (u->is_MemBar()) { 3180 if (u->as_MemBar()->trailing_load_store()) { 3181 assert(u->Opcode() == Op_MemBarAcquire, ""); 3182 assert(trailing == nullptr, "only one"); 3183 trailing = u->as_MemBar(); 3184 #ifdef ASSERT 3185 Node* leading = trailing->leading_membar(); 3186 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 3187 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair"); 3188 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair"); 3189 #endif 3190 } else { 3191 assert(u->as_MemBar()->standalone(), "wrong barrier kind"); 3192 } 3193 } 3194 } 3195 3196 return trailing; 3197 } 3198 3199 uint LoadStoreNode::size_of() const { return sizeof(*this); } 3200 3201 //============================================================================= 3202 //----------------------------------LoadStoreConditionalNode-------------------- 3203 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) { 3204 init_req(ExpectedIn, ex ); 3205 } 3206 3207 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const { 3208 // Either input is TOP ==> the result is TOP 3209 const Type* t = phase->type(in(ExpectedIn)); 3210 if (t == Type::TOP) { 3211 return Type::TOP; 3212 } 3213 return LoadStoreNode::Value(phase); 3214 } 3215 3216 //============================================================================= 3217 //-------------------------------adr_type-------------------------------------- 3218 const TypePtr* ClearArrayNode::adr_type() const { 3219 Node *adr = in(3); 3220 if (adr == nullptr) return nullptr; // node is dead 3221 return MemNode::calculate_adr_type(adr->bottom_type()); 3222 } 3223 3224 //------------------------------match_edge------------------------------------- 3225 // Do we Match on this edge index or not? Do not match memory 3226 uint ClearArrayNode::match_edge(uint idx) const { 3227 return idx > 1; 3228 } 3229 3230 //------------------------------Identity--------------------------------------- 3231 // Clearing a zero length array does nothing 3232 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 3233 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 3234 } 3235 3236 //------------------------------Idealize--------------------------------------- 3237 // Clearing a short array is faster with stores 3238 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3239 // Already know this is a large node, do not try to ideal it 3240 if (_is_large) return nullptr; 3241 3242 const int unit = BytesPerLong; 3243 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 3244 if (!t) return nullptr; 3245 if (!t->is_con()) return nullptr; 3246 intptr_t raw_count = t->get_con(); 3247 intptr_t size = raw_count; 3248 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 3249 // Clearing nothing uses the Identity call. 3250 // Negative clears are possible on dead ClearArrays 3251 // (see jck test stmt114.stmt11402.val). 3252 if (size <= 0 || size % unit != 0) return nullptr; 3253 intptr_t count = size / unit; 3254 // Length too long; communicate this to matchers and assemblers. 3255 // Assemblers are responsible to produce fast hardware clears for it. 3256 if (size > InitArrayShortSize) { 3257 return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true); 3258 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) { 3259 return nullptr; 3260 } 3261 if (!IdealizeClearArrayNode) return nullptr; 3262 Node *mem = in(1); 3263 if( phase->type(mem)==Type::TOP ) return nullptr; 3264 Node *adr = in(3); 3265 const Type* at = phase->type(adr); 3266 if( at==Type::TOP ) return nullptr; 3267 const TypePtr* atp = at->isa_ptr(); 3268 // adjust atp to be the correct array element address type 3269 if (atp == nullptr) atp = TypePtr::BOTTOM; 3270 else atp = atp->add_offset(Type::OffsetBot); 3271 // Get base for derived pointer purposes 3272 if( adr->Opcode() != Op_AddP ) Unimplemented(); 3273 Node *base = adr->in(1); 3274 3275 Node *val = in(4); 3276 Node *off = phase->MakeConX(BytesPerLong); 3277 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false); 3278 count--; 3279 while( count-- ) { 3280 mem = phase->transform(mem); 3281 adr = phase->transform(new AddPNode(base,adr,off)); 3282 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false); 3283 } 3284 return mem; 3285 } 3286 3287 //----------------------------step_through---------------------------------- 3288 // Return allocation input memory edge if it is different instance 3289 // or itself if it is the one we are looking for. 3290 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) { 3291 Node* n = *np; 3292 assert(n->is_ClearArray(), "sanity"); 3293 intptr_t offset; 3294 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 3295 // This method is called only before Allocate nodes are expanded 3296 // during macro nodes expansion. Before that ClearArray nodes are 3297 // only generated in PhaseMacroExpand::generate_arraycopy() (before 3298 // Allocate nodes are expanded) which follows allocations. 3299 assert(alloc != nullptr, "should have allocation"); 3300 if (alloc->_idx == instance_id) { 3301 // Can not bypass initialization of the instance we are looking for. 3302 return false; 3303 } 3304 // Otherwise skip it. 3305 InitializeNode* init = alloc->initialization(); 3306 if (init != nullptr) 3307 *np = init->in(TypeFunc::Memory); 3308 else 3309 *np = alloc->in(TypeFunc::Memory); 3310 return true; 3311 } 3312 3313 //----------------------------clear_memory------------------------------------- 3314 // Generate code to initialize object storage to zero. 3315 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3316 Node* val, 3317 Node* raw_val, 3318 intptr_t start_offset, 3319 Node* end_offset, 3320 PhaseGVN* phase) { 3321 intptr_t offset = start_offset; 3322 3323 int unit = BytesPerLong; 3324 if ((offset % unit) != 0) { 3325 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 3326 adr = phase->transform(adr); 3327 const TypePtr* atp = TypeRawPtr::BOTTOM; 3328 if (val != nullptr) { 3329 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop"); 3330 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered); 3331 } else { 3332 assert(raw_val == nullptr, "val may not be null"); 3333 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3334 } 3335 mem = phase->transform(mem); 3336 offset += BytesPerInt; 3337 } 3338 assert((offset % unit) == 0, ""); 3339 3340 // Initialize the remaining stuff, if any, with a ClearArray. 3341 return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase); 3342 } 3343 3344 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3345 Node* raw_val, 3346 Node* start_offset, 3347 Node* end_offset, 3348 PhaseGVN* phase) { 3349 if (start_offset == end_offset) { 3350 // nothing to do 3351 return mem; 3352 } 3353 3354 int unit = BytesPerLong; 3355 Node* zbase = start_offset; 3356 Node* zend = end_offset; 3357 3358 // Scale to the unit required by the CPU: 3359 if (!Matcher::init_array_count_is_in_bytes) { 3360 Node* shift = phase->intcon(exact_log2(unit)); 3361 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 3362 zend = phase->transform(new URShiftXNode(zend, shift) ); 3363 } 3364 3365 // Bulk clear double-words 3366 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 3367 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 3368 if (raw_val == nullptr) { 3369 raw_val = phase->MakeConX(0); 3370 } 3371 mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false); 3372 return phase->transform(mem); 3373 } 3374 3375 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3376 Node* val, 3377 Node* raw_val, 3378 intptr_t start_offset, 3379 intptr_t end_offset, 3380 PhaseGVN* phase) { 3381 if (start_offset == end_offset) { 3382 // nothing to do 3383 return mem; 3384 } 3385 3386 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 3387 intptr_t done_offset = end_offset; 3388 if ((done_offset % BytesPerLong) != 0) { 3389 done_offset -= BytesPerInt; 3390 } 3391 if (done_offset > start_offset) { 3392 mem = clear_memory(ctl, mem, dest, val, raw_val, 3393 start_offset, phase->MakeConX(done_offset), phase); 3394 } 3395 if (done_offset < end_offset) { // emit the final 32-bit store 3396 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 3397 adr = phase->transform(adr); 3398 const TypePtr* atp = TypeRawPtr::BOTTOM; 3399 if (val != nullptr) { 3400 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop"); 3401 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered); 3402 } else { 3403 assert(raw_val == nullptr, "val may not be null"); 3404 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3405 } 3406 mem = phase->transform(mem); 3407 done_offset += BytesPerInt; 3408 } 3409 assert(done_offset == end_offset, ""); 3410 return mem; 3411 } 3412 3413 //============================================================================= 3414 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 3415 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)), 3416 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone) 3417 #ifdef ASSERT 3418 , _pair_idx(0) 3419 #endif 3420 { 3421 init_class_id(Class_MemBar); 3422 Node* top = C->top(); 3423 init_req(TypeFunc::I_O,top); 3424 init_req(TypeFunc::FramePtr,top); 3425 init_req(TypeFunc::ReturnAdr,top); 3426 if (precedent != nullptr) 3427 init_req(TypeFunc::Parms, precedent); 3428 } 3429 3430 //------------------------------cmp-------------------------------------------- 3431 uint MemBarNode::hash() const { return NO_HASH; } 3432 bool MemBarNode::cmp( const Node &n ) const { 3433 return (&n == this); // Always fail except on self 3434 } 3435 3436 //------------------------------make------------------------------------------- 3437 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 3438 switch (opcode) { 3439 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 3440 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 3441 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 3442 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 3443 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 3444 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn); 3445 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 3446 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 3447 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 3448 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 3449 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 3450 case Op_Initialize: return new InitializeNode(C, atp, pn); 3451 default: ShouldNotReachHere(); return nullptr; 3452 } 3453 } 3454 3455 void MemBarNode::remove(PhaseIterGVN *igvn) { 3456 if (outcnt() != 2) { 3457 assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory"); 3458 assert(outcnt() == 1, "Only control then"); 3459 } 3460 if (trailing_store() || trailing_load_store()) { 3461 MemBarNode* leading = leading_membar(); 3462 if (leading != nullptr) { 3463 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars"); 3464 leading->remove(igvn); 3465 } 3466 } 3467 if (proj_out_or_null(TypeFunc::Memory) != nullptr) { 3468 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 3469 } 3470 if (proj_out_or_null(TypeFunc::Control) != nullptr) { 3471 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 3472 } 3473 } 3474 3475 //------------------------------Ideal------------------------------------------ 3476 // Return a node which is more "ideal" than the current node. Strip out 3477 // control copies 3478 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3479 if (remove_dead_region(phase, can_reshape)) return this; 3480 // Don't bother trying to transform a dead node 3481 if (in(0) && in(0)->is_top()) { 3482 return nullptr; 3483 } 3484 3485 bool progress = false; 3486 // Eliminate volatile MemBars for scalar replaced objects. 3487 if (can_reshape && req() == (Precedent+1)) { 3488 bool eliminate = false; 3489 int opc = Opcode(); 3490 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 3491 // Volatile field loads and stores. 3492 Node* my_mem = in(MemBarNode::Precedent); 3493 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 3494 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 3495 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 3496 // replace this Precedent (decodeN) with the Load instead. 3497 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 3498 Node* load_node = my_mem->in(1); 3499 set_req(MemBarNode::Precedent, load_node); 3500 phase->is_IterGVN()->_worklist.push(my_mem); 3501 my_mem = load_node; 3502 } else { 3503 assert(my_mem->unique_out() == this, "sanity"); 3504 del_req(Precedent); 3505 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 3506 my_mem = nullptr; 3507 } 3508 progress = true; 3509 } 3510 if (my_mem != nullptr && my_mem->is_Mem()) { 3511 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 3512 // Check for scalar replaced object reference. 3513 if( t_oop != nullptr && t_oop->is_known_instance_field() && 3514 t_oop->offset() != Type::OffsetBot && 3515 t_oop->offset() != Type::OffsetTop) { 3516 eliminate = true; 3517 } 3518 } 3519 } else if (opc == Op_MemBarRelease) { 3520 // Final field stores. 3521 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent)); 3522 if ((alloc != nullptr) && alloc->is_Allocate() && 3523 alloc->as_Allocate()->does_not_escape_thread()) { 3524 // The allocated object does not escape. 3525 eliminate = true; 3526 } 3527 } 3528 if (eliminate) { 3529 // Replace MemBar projections by its inputs. 3530 PhaseIterGVN* igvn = phase->is_IterGVN(); 3531 remove(igvn); 3532 // Must return either the original node (now dead) or a new node 3533 // (Do not return a top here, since that would break the uniqueness of top.) 3534 return new ConINode(TypeInt::ZERO); 3535 } 3536 } 3537 return progress ? this : nullptr; 3538 } 3539 3540 //------------------------------Value------------------------------------------ 3541 const Type* MemBarNode::Value(PhaseGVN* phase) const { 3542 if( !in(0) ) return Type::TOP; 3543 if( phase->type(in(0)) == Type::TOP ) 3544 return Type::TOP; 3545 return TypeTuple::MEMBAR; 3546 } 3547 3548 //------------------------------match------------------------------------------ 3549 // Construct projections for memory. 3550 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) { 3551 switch (proj->_con) { 3552 case TypeFunc::Control: 3553 case TypeFunc::Memory: 3554 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3555 } 3556 ShouldNotReachHere(); 3557 return nullptr; 3558 } 3559 3560 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3561 trailing->_kind = TrailingStore; 3562 leading->_kind = LeadingStore; 3563 #ifdef ASSERT 3564 trailing->_pair_idx = leading->_idx; 3565 leading->_pair_idx = leading->_idx; 3566 #endif 3567 } 3568 3569 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3570 trailing->_kind = TrailingLoadStore; 3571 leading->_kind = LeadingLoadStore; 3572 #ifdef ASSERT 3573 trailing->_pair_idx = leading->_idx; 3574 leading->_pair_idx = leading->_idx; 3575 #endif 3576 } 3577 3578 MemBarNode* MemBarNode::trailing_membar() const { 3579 ResourceMark rm; 3580 Node* trailing = (Node*)this; 3581 VectorSet seen; 3582 Node_Stack multis(0); 3583 do { 3584 Node* c = trailing; 3585 uint i = 0; 3586 do { 3587 trailing = nullptr; 3588 for (; i < c->outcnt(); i++) { 3589 Node* next = c->raw_out(i); 3590 if (next != c && next->is_CFG()) { 3591 if (c->is_MultiBranch()) { 3592 if (multis.node() == c) { 3593 multis.set_index(i+1); 3594 } else { 3595 multis.push(c, i+1); 3596 } 3597 } 3598 trailing = next; 3599 break; 3600 } 3601 } 3602 if (trailing != nullptr && !seen.test_set(trailing->_idx)) { 3603 break; 3604 } 3605 while (multis.size() > 0) { 3606 c = multis.node(); 3607 i = multis.index(); 3608 if (i < c->req()) { 3609 break; 3610 } 3611 multis.pop(); 3612 } 3613 } while (multis.size() > 0); 3614 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing()); 3615 3616 MemBarNode* mb = trailing->as_MemBar(); 3617 assert((mb->_kind == TrailingStore && _kind == LeadingStore) || 3618 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar"); 3619 assert(mb->_pair_idx == _pair_idx, "bad trailing membar"); 3620 return mb; 3621 } 3622 3623 MemBarNode* MemBarNode::leading_membar() const { 3624 ResourceMark rm; 3625 VectorSet seen; 3626 Node_Stack regions(0); 3627 Node* leading = in(0); 3628 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) { 3629 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) { 3630 leading = nullptr; 3631 while (regions.size() > 0 && leading == nullptr) { 3632 Node* r = regions.node(); 3633 uint i = regions.index(); 3634 if (i < r->req()) { 3635 leading = r->in(i); 3636 regions.set_index(i+1); 3637 } else { 3638 regions.pop(); 3639 } 3640 } 3641 if (leading == nullptr) { 3642 assert(regions.size() == 0, "all paths should have been tried"); 3643 return nullptr; 3644 } 3645 } 3646 if (leading->is_Region()) { 3647 regions.push(leading, 2); 3648 leading = leading->in(1); 3649 } else { 3650 leading = leading->in(0); 3651 } 3652 } 3653 #ifdef ASSERT 3654 Unique_Node_List wq; 3655 wq.push((Node*)this); 3656 uint found = 0; 3657 for (uint i = 0; i < wq.size(); i++) { 3658 Node* n = wq.at(i); 3659 if (n->is_Region()) { 3660 for (uint j = 1; j < n->req(); j++) { 3661 Node* in = n->in(j); 3662 if (in != nullptr && !in->is_top()) { 3663 wq.push(in); 3664 } 3665 } 3666 } else { 3667 if (n->is_MemBar() && n->as_MemBar()->leading()) { 3668 assert(n == leading, "consistency check failed"); 3669 found++; 3670 } else { 3671 Node* in = n->in(0); 3672 if (in != nullptr && !in->is_top()) { 3673 wq.push(in); 3674 } 3675 } 3676 } 3677 } 3678 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed"); 3679 #endif 3680 if (leading == nullptr) { 3681 return nullptr; 3682 } 3683 MemBarNode* mb = leading->as_MemBar(); 3684 assert((mb->_kind == LeadingStore && _kind == TrailingStore) || 3685 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar"); 3686 assert(mb->_pair_idx == _pair_idx, "bad leading membar"); 3687 return mb; 3688 } 3689 3690 3691 //===========================InitializeNode==================================== 3692 // SUMMARY: 3693 // This node acts as a memory barrier on raw memory, after some raw stores. 3694 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3695 // The Initialize can 'capture' suitably constrained stores as raw inits. 3696 // It can coalesce related raw stores into larger units (called 'tiles'). 3697 // It can avoid zeroing new storage for memory units which have raw inits. 3698 // At macro-expansion, it is marked 'complete', and does not optimize further. 3699 // 3700 // EXAMPLE: 3701 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3702 // ctl = incoming control; mem* = incoming memory 3703 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3704 // First allocate uninitialized memory and fill in the header: 3705 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3706 // ctl := alloc.Control; mem* := alloc.Memory* 3707 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3708 // Then initialize to zero the non-header parts of the raw memory block: 3709 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3710 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3711 // After the initialize node executes, the object is ready for service: 3712 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3713 // Suppose its body is immediately initialized as {1,2}: 3714 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3715 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3716 // mem.SLICE(#short[*]) := store2 3717 // 3718 // DETAILS: 3719 // An InitializeNode collects and isolates object initialization after 3720 // an AllocateNode and before the next possible safepoint. As a 3721 // memory barrier (MemBarNode), it keeps critical stores from drifting 3722 // down past any safepoint or any publication of the allocation. 3723 // Before this barrier, a newly-allocated object may have uninitialized bits. 3724 // After this barrier, it may be treated as a real oop, and GC is allowed. 3725 // 3726 // The semantics of the InitializeNode include an implicit zeroing of 3727 // the new object from object header to the end of the object. 3728 // (The object header and end are determined by the AllocateNode.) 3729 // 3730 // Certain stores may be added as direct inputs to the InitializeNode. 3731 // These stores must update raw memory, and they must be to addresses 3732 // derived from the raw address produced by AllocateNode, and with 3733 // a constant offset. They must be ordered by increasing offset. 3734 // The first one is at in(RawStores), the last at in(req()-1). 3735 // Unlike most memory operations, they are not linked in a chain, 3736 // but are displayed in parallel as users of the rawmem output of 3737 // the allocation. 3738 // 3739 // (See comments in InitializeNode::capture_store, which continue 3740 // the example given above.) 3741 // 3742 // When the associated Allocate is macro-expanded, the InitializeNode 3743 // may be rewritten to optimize collected stores. A ClearArrayNode 3744 // may also be created at that point to represent any required zeroing. 3745 // The InitializeNode is then marked 'complete', prohibiting further 3746 // capturing of nearby memory operations. 3747 // 3748 // During macro-expansion, all captured initializations which store 3749 // constant values of 32 bits or smaller are coalesced (if advantageous) 3750 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3751 // initialized in fewer memory operations. Memory words which are 3752 // covered by neither tiles nor non-constant stores are pre-zeroed 3753 // by explicit stores of zero. (The code shape happens to do all 3754 // zeroing first, then all other stores, with both sequences occurring 3755 // in order of ascending offsets.) 3756 // 3757 // Alternatively, code may be inserted between an AllocateNode and its 3758 // InitializeNode, to perform arbitrary initialization of the new object. 3759 // E.g., the object copying intrinsics insert complex data transfers here. 3760 // The initialization must then be marked as 'complete' disable the 3761 // built-in zeroing semantics and the collection of initializing stores. 3762 // 3763 // While an InitializeNode is incomplete, reads from the memory state 3764 // produced by it are optimizable if they match the control edge and 3765 // new oop address associated with the allocation/initialization. 3766 // They return a stored value (if the offset matches) or else zero. 3767 // A write to the memory state, if it matches control and address, 3768 // and if it is to a constant offset, may be 'captured' by the 3769 // InitializeNode. It is cloned as a raw memory operation and rewired 3770 // inside the initialization, to the raw oop produced by the allocation. 3771 // Operations on addresses which are provably distinct (e.g., to 3772 // other AllocateNodes) are allowed to bypass the initialization. 3773 // 3774 // The effect of all this is to consolidate object initialization 3775 // (both arrays and non-arrays, both piecewise and bulk) into a 3776 // single location, where it can be optimized as a unit. 3777 // 3778 // Only stores with an offset less than TrackedInitializationLimit words 3779 // will be considered for capture by an InitializeNode. This puts a 3780 // reasonable limit on the complexity of optimized initializations. 3781 3782 //---------------------------InitializeNode------------------------------------ 3783 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3784 : MemBarNode(C, adr_type, rawoop), 3785 _is_complete(Incomplete), _does_not_escape(false) 3786 { 3787 init_class_id(Class_Initialize); 3788 3789 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3790 assert(in(RawAddress) == rawoop, "proper init"); 3791 // Note: allocation() can be null, for secondary initialization barriers 3792 } 3793 3794 // Since this node is not matched, it will be processed by the 3795 // register allocator. Declare that there are no constraints 3796 // on the allocation of the RawAddress edge. 3797 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3798 // This edge should be set to top, by the set_complete. But be conservative. 3799 if (idx == InitializeNode::RawAddress) 3800 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3801 return RegMask::Empty; 3802 } 3803 3804 Node* InitializeNode::memory(uint alias_idx) { 3805 Node* mem = in(Memory); 3806 if (mem->is_MergeMem()) { 3807 return mem->as_MergeMem()->memory_at(alias_idx); 3808 } else { 3809 // incoming raw memory is not split 3810 return mem; 3811 } 3812 } 3813 3814 bool InitializeNode::is_non_zero() { 3815 if (is_complete()) return false; 3816 remove_extra_zeroes(); 3817 return (req() > RawStores); 3818 } 3819 3820 void InitializeNode::set_complete(PhaseGVN* phase) { 3821 assert(!is_complete(), "caller responsibility"); 3822 _is_complete = Complete; 3823 3824 // After this node is complete, it contains a bunch of 3825 // raw-memory initializations. There is no need for 3826 // it to have anything to do with non-raw memory effects. 3827 // Therefore, tell all non-raw users to re-optimize themselves, 3828 // after skipping the memory effects of this initialization. 3829 PhaseIterGVN* igvn = phase->is_IterGVN(); 3830 if (igvn) igvn->add_users_to_worklist(this); 3831 } 3832 3833 // convenience function 3834 // return false if the init contains any stores already 3835 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3836 InitializeNode* init = initialization(); 3837 if (init == nullptr || init->is_complete()) { 3838 return false; 3839 } 3840 init->remove_extra_zeroes(); 3841 // for now, if this allocation has already collected any inits, bail: 3842 if (init->is_non_zero()) return false; 3843 init->set_complete(phase); 3844 return true; 3845 } 3846 3847 void InitializeNode::remove_extra_zeroes() { 3848 if (req() == RawStores) return; 3849 Node* zmem = zero_memory(); 3850 uint fill = RawStores; 3851 for (uint i = fill; i < req(); i++) { 3852 Node* n = in(i); 3853 if (n->is_top() || n == zmem) continue; // skip 3854 if (fill < i) set_req(fill, n); // compact 3855 ++fill; 3856 } 3857 // delete any empty spaces created: 3858 while (fill < req()) { 3859 del_req(fill); 3860 } 3861 } 3862 3863 // Helper for remembering which stores go with which offsets. 3864 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) { 3865 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3866 intptr_t offset = -1; 3867 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3868 phase, offset); 3869 if (base == nullptr) return -1; // something is dead, 3870 if (offset < 0) return -1; // dead, dead 3871 return offset; 3872 } 3873 3874 // Helper for proving that an initialization expression is 3875 // "simple enough" to be folded into an object initialization. 3876 // Attempts to prove that a store's initial value 'n' can be captured 3877 // within the initialization without creating a vicious cycle, such as: 3878 // { Foo p = new Foo(); p.next = p; } 3879 // True for constants and parameters and small combinations thereof. 3880 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) { 3881 ResourceMark rm; 3882 Unique_Node_List worklist; 3883 worklist.push(value); 3884 3885 uint complexity_limit = 20; 3886 for (uint j = 0; j < worklist.size(); j++) { 3887 if (j >= complexity_limit) { 3888 return false; // Bail out if processed too many nodes 3889 } 3890 3891 Node* n = worklist.at(j); 3892 if (n == nullptr) continue; // (can this really happen?) 3893 if (n->is_Proj()) n = n->in(0); 3894 if (n == this) return false; // found a cycle 3895 if (n->is_Con()) continue; 3896 if (n->is_Start()) continue; // params, etc., are OK 3897 if (n->is_Root()) continue; // even better 3898 3899 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation 3900 if (n->is_CFG() && phase->is_dominator(n, allocation())) { 3901 continue; 3902 } 3903 3904 Node* ctl = n->in(0); 3905 if (ctl != nullptr && !ctl->is_top()) { 3906 if (ctl->is_Proj()) ctl = ctl->in(0); 3907 if (ctl == this) return false; 3908 3909 // If we already know that the enclosing memory op is pinned right after 3910 // the init, then any control flow that the store has picked up 3911 // must have preceded the init, or else be equal to the init. 3912 // Even after loop optimizations (which might change control edges) 3913 // a store is never pinned *before* the availability of its inputs. 3914 if (!MemNode::all_controls_dominate(n, this)) 3915 return false; // failed to prove a good control 3916 } 3917 3918 // Check data edges for possible dependencies on 'this'. 3919 for (uint i = 1; i < n->req(); i++) { 3920 Node* m = n->in(i); 3921 if (m == nullptr || m == n || m->is_top()) continue; 3922 3923 // Only process data inputs once 3924 worklist.push(m); 3925 } 3926 } 3927 3928 return true; 3929 } 3930 3931 // Here are all the checks a Store must pass before it can be moved into 3932 // an initialization. Returns zero if a check fails. 3933 // On success, returns the (constant) offset to which the store applies, 3934 // within the initialized memory. 3935 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) { 3936 const int FAIL = 0; 3937 if (st->req() != MemNode::ValueIn + 1) 3938 return FAIL; // an inscrutable StoreNode (card mark?) 3939 Node* ctl = st->in(MemNode::Control); 3940 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this)) 3941 return FAIL; // must be unconditional after the initialization 3942 Node* mem = st->in(MemNode::Memory); 3943 if (!(mem->is_Proj() && mem->in(0) == this)) 3944 return FAIL; // must not be preceded by other stores 3945 Node* adr = st->in(MemNode::Address); 3946 intptr_t offset; 3947 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3948 if (alloc == nullptr) 3949 return FAIL; // inscrutable address 3950 if (alloc != allocation()) 3951 return FAIL; // wrong allocation! (store needs to float up) 3952 int size_in_bytes = st->memory_size(); 3953 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) { 3954 return FAIL; // mismatched access 3955 } 3956 Node* val = st->in(MemNode::ValueIn); 3957 3958 if (!detect_init_independence(val, phase)) 3959 return FAIL; // stored value must be 'simple enough' 3960 3961 // The Store can be captured only if nothing after the allocation 3962 // and before the Store is using the memory location that the store 3963 // overwrites. 3964 bool failed = false; 3965 // If is_complete_with_arraycopy() is true the shape of the graph is 3966 // well defined and is safe so no need for extra checks. 3967 if (!is_complete_with_arraycopy()) { 3968 // We are going to look at each use of the memory state following 3969 // the allocation to make sure nothing reads the memory that the 3970 // Store writes. 3971 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3972 int alias_idx = phase->C->get_alias_index(t_adr); 3973 ResourceMark rm; 3974 Unique_Node_List mems; 3975 mems.push(mem); 3976 Node* unique_merge = nullptr; 3977 for (uint next = 0; next < mems.size(); ++next) { 3978 Node *m = mems.at(next); 3979 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3980 Node *n = m->fast_out(j); 3981 if (n->outcnt() == 0) { 3982 continue; 3983 } 3984 if (n == st) { 3985 continue; 3986 } else if (n->in(0) != nullptr && n->in(0) != ctl) { 3987 // If the control of this use is different from the control 3988 // of the Store which is right after the InitializeNode then 3989 // this node cannot be between the InitializeNode and the 3990 // Store. 3991 continue; 3992 } else if (n->is_MergeMem()) { 3993 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3994 // We can hit a MergeMemNode (that will likely go away 3995 // later) that is a direct use of the memory state 3996 // following the InitializeNode on the same slice as the 3997 // store node that we'd like to capture. We need to check 3998 // the uses of the MergeMemNode. 3999 mems.push(n); 4000 } 4001 } else if (n->is_Mem()) { 4002 Node* other_adr = n->in(MemNode::Address); 4003 if (other_adr == adr) { 4004 failed = true; 4005 break; 4006 } else { 4007 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 4008 if (other_t_adr != nullptr) { 4009 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 4010 if (other_alias_idx == alias_idx) { 4011 // A load from the same memory slice as the store right 4012 // after the InitializeNode. We check the control of the 4013 // object/array that is loaded from. If it's the same as 4014 // the store control then we cannot capture the store. 4015 assert(!n->is_Store(), "2 stores to same slice on same control?"); 4016 Node* base = other_adr; 4017 if (base->is_Phi()) { 4018 // In rare case, base may be a PhiNode and it may read 4019 // the same memory slice between InitializeNode and store. 4020 failed = true; 4021 break; 4022 } 4023 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 4024 base = base->in(AddPNode::Base); 4025 if (base != nullptr) { 4026 base = base->uncast(); 4027 if (base->is_Proj() && base->in(0) == alloc) { 4028 failed = true; 4029 break; 4030 } 4031 } 4032 } 4033 } 4034 } 4035 } else { 4036 failed = true; 4037 break; 4038 } 4039 } 4040 } 4041 } 4042 if (failed) { 4043 if (!can_reshape) { 4044 // We decided we couldn't capture the store during parsing. We 4045 // should try again during the next IGVN once the graph is 4046 // cleaner. 4047 phase->C->record_for_igvn(st); 4048 } 4049 return FAIL; 4050 } 4051 4052 return offset; // success 4053 } 4054 4055 // Find the captured store in(i) which corresponds to the range 4056 // [start..start+size) in the initialized object. 4057 // If there is one, return its index i. If there isn't, return the 4058 // negative of the index where it should be inserted. 4059 // Return 0 if the queried range overlaps an initialization boundary 4060 // or if dead code is encountered. 4061 // If size_in_bytes is zero, do not bother with overlap checks. 4062 int InitializeNode::captured_store_insertion_point(intptr_t start, 4063 int size_in_bytes, 4064 PhaseValues* phase) { 4065 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize); 4066 4067 if (is_complete()) 4068 return FAIL; // arraycopy got here first; punt 4069 4070 assert(allocation() != nullptr, "must be present"); 4071 4072 // no negatives, no header fields: 4073 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 4074 4075 // after a certain size, we bail out on tracking all the stores: 4076 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 4077 if (start >= ti_limit) return FAIL; 4078 4079 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 4080 if (i >= limit) return -(int)i; // not found; here is where to put it 4081 4082 Node* st = in(i); 4083 intptr_t st_off = get_store_offset(st, phase); 4084 if (st_off < 0) { 4085 if (st != zero_memory()) { 4086 return FAIL; // bail out if there is dead garbage 4087 } 4088 } else if (st_off > start) { 4089 // ...we are done, since stores are ordered 4090 if (st_off < start + size_in_bytes) { 4091 return FAIL; // the next store overlaps 4092 } 4093 return -(int)i; // not found; here is where to put it 4094 } else if (st_off < start) { 4095 assert(st->as_Store()->memory_size() <= MAX_STORE, ""); 4096 if (size_in_bytes != 0 && 4097 start < st_off + MAX_STORE && 4098 start < st_off + st->as_Store()->memory_size()) { 4099 return FAIL; // the previous store overlaps 4100 } 4101 } else { 4102 if (size_in_bytes != 0 && 4103 st->as_Store()->memory_size() != size_in_bytes) { 4104 return FAIL; // mismatched store size 4105 } 4106 return i; 4107 } 4108 4109 ++i; 4110 } 4111 } 4112 4113 // Look for a captured store which initializes at the offset 'start' 4114 // with the given size. If there is no such store, and no other 4115 // initialization interferes, then return zero_memory (the memory 4116 // projection of the AllocateNode). 4117 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 4118 PhaseValues* phase) { 4119 assert(stores_are_sane(phase), ""); 4120 int i = captured_store_insertion_point(start, size_in_bytes, phase); 4121 if (i == 0) { 4122 return nullptr; // something is dead 4123 } else if (i < 0) { 4124 return zero_memory(); // just primordial zero bits here 4125 } else { 4126 Node* st = in(i); // here is the store at this position 4127 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 4128 return st; 4129 } 4130 } 4131 4132 // Create, as a raw pointer, an address within my new object at 'offset'. 4133 Node* InitializeNode::make_raw_address(intptr_t offset, 4134 PhaseGVN* phase) { 4135 Node* addr = in(RawAddress); 4136 if (offset != 0) { 4137 Compile* C = phase->C; 4138 addr = phase->transform( new AddPNode(C->top(), addr, 4139 phase->MakeConX(offset)) ); 4140 } 4141 return addr; 4142 } 4143 4144 // Clone the given store, converting it into a raw store 4145 // initializing a field or element of my new object. 4146 // Caller is responsible for retiring the original store, 4147 // with subsume_node or the like. 4148 // 4149 // From the example above InitializeNode::InitializeNode, 4150 // here are the old stores to be captured: 4151 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 4152 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 4153 // 4154 // Here is the changed code; note the extra edges on init: 4155 // alloc = (Allocate ...) 4156 // rawoop = alloc.RawAddress 4157 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 4158 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 4159 // init = (Initialize alloc.Control alloc.Memory rawoop 4160 // rawstore1 rawstore2) 4161 // 4162 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 4163 PhaseGVN* phase, bool can_reshape) { 4164 assert(stores_are_sane(phase), ""); 4165 4166 if (start < 0) return nullptr; 4167 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 4168 4169 Compile* C = phase->C; 4170 int size_in_bytes = st->memory_size(); 4171 int i = captured_store_insertion_point(start, size_in_bytes, phase); 4172 if (i == 0) return nullptr; // bail out 4173 Node* prev_mem = nullptr; // raw memory for the captured store 4174 if (i > 0) { 4175 prev_mem = in(i); // there is a pre-existing store under this one 4176 set_req(i, C->top()); // temporarily disconnect it 4177 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 4178 } else { 4179 i = -i; // no pre-existing store 4180 prev_mem = zero_memory(); // a slice of the newly allocated object 4181 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 4182 set_req(--i, C->top()); // reuse this edge; it has been folded away 4183 else 4184 ins_req(i, C->top()); // build a new edge 4185 } 4186 Node* new_st = st->clone(); 4187 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4188 new_st->set_req(MemNode::Control, in(Control)); 4189 new_st->set_req(MemNode::Memory, prev_mem); 4190 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 4191 bs->eliminate_gc_barrier_data(new_st); 4192 new_st = phase->transform(new_st); 4193 4194 // At this point, new_st might have swallowed a pre-existing store 4195 // at the same offset, or perhaps new_st might have disappeared, 4196 // if it redundantly stored the same value (or zero to fresh memory). 4197 4198 // In any case, wire it in: 4199 PhaseIterGVN* igvn = phase->is_IterGVN(); 4200 if (igvn) { 4201 igvn->rehash_node_delayed(this); 4202 } 4203 set_req(i, new_st); 4204 4205 // The caller may now kill the old guy. 4206 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 4207 assert(check_st == new_st || check_st == nullptr, "must be findable"); 4208 assert(!is_complete(), ""); 4209 return new_st; 4210 } 4211 4212 static bool store_constant(jlong* tiles, int num_tiles, 4213 intptr_t st_off, int st_size, 4214 jlong con) { 4215 if ((st_off & (st_size-1)) != 0) 4216 return false; // strange store offset (assume size==2**N) 4217 address addr = (address)tiles + st_off; 4218 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 4219 switch (st_size) { 4220 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 4221 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 4222 case sizeof(jint): *(jint*) addr = (jint) con; break; 4223 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 4224 default: return false; // strange store size (detect size!=2**N here) 4225 } 4226 return true; // return success to caller 4227 } 4228 4229 // Coalesce subword constants into int constants and possibly 4230 // into long constants. The goal, if the CPU permits, 4231 // is to initialize the object with a small number of 64-bit tiles. 4232 // Also, convert floating-point constants to bit patterns. 4233 // Non-constants are not relevant to this pass. 4234 // 4235 // In terms of the running example on InitializeNode::InitializeNode 4236 // and InitializeNode::capture_store, here is the transformation 4237 // of rawstore1 and rawstore2 into rawstore12: 4238 // alloc = (Allocate ...) 4239 // rawoop = alloc.RawAddress 4240 // tile12 = 0x00010002 4241 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 4242 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 4243 // 4244 void 4245 InitializeNode::coalesce_subword_stores(intptr_t header_size, 4246 Node* size_in_bytes, 4247 PhaseGVN* phase) { 4248 Compile* C = phase->C; 4249 4250 assert(stores_are_sane(phase), ""); 4251 // Note: After this pass, they are not completely sane, 4252 // since there may be some overlaps. 4253 4254 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 4255 4256 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 4257 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 4258 size_limit = MIN2(size_limit, ti_limit); 4259 size_limit = align_up(size_limit, BytesPerLong); 4260 int num_tiles = size_limit / BytesPerLong; 4261 4262 // allocate space for the tile map: 4263 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 4264 jlong tiles_buf[small_len]; 4265 Node* nodes_buf[small_len]; 4266 jlong inits_buf[small_len]; 4267 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 4268 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 4269 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 4270 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 4271 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 4272 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 4273 // tiles: exact bitwise model of all primitive constants 4274 // nodes: last constant-storing node subsumed into the tiles model 4275 // inits: which bytes (in each tile) are touched by any initializations 4276 4277 //// Pass A: Fill in the tile model with any relevant stores. 4278 4279 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 4280 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 4281 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 4282 Node* zmem = zero_memory(); // initially zero memory state 4283 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4284 Node* st = in(i); 4285 intptr_t st_off = get_store_offset(st, phase); 4286 4287 // Figure out the store's offset and constant value: 4288 if (st_off < header_size) continue; //skip (ignore header) 4289 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 4290 int st_size = st->as_Store()->memory_size(); 4291 if (st_off + st_size > size_limit) break; 4292 4293 // Record which bytes are touched, whether by constant or not. 4294 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 4295 continue; // skip (strange store size) 4296 4297 const Type* val = phase->type(st->in(MemNode::ValueIn)); 4298 if (!val->singleton()) continue; //skip (non-con store) 4299 BasicType type = val->basic_type(); 4300 4301 jlong con = 0; 4302 switch (type) { 4303 case T_INT: con = val->is_int()->get_con(); break; 4304 case T_LONG: con = val->is_long()->get_con(); break; 4305 case T_FLOAT: con = jint_cast(val->getf()); break; 4306 case T_DOUBLE: con = jlong_cast(val->getd()); break; 4307 default: continue; //skip (odd store type) 4308 } 4309 4310 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 4311 st->Opcode() == Op_StoreL) { 4312 continue; // This StoreL is already optimal. 4313 } 4314 4315 // Store down the constant. 4316 store_constant(tiles, num_tiles, st_off, st_size, con); 4317 4318 intptr_t j = st_off >> LogBytesPerLong; 4319 4320 if (type == T_INT && st_size == BytesPerInt 4321 && (st_off & BytesPerInt) == BytesPerInt) { 4322 jlong lcon = tiles[j]; 4323 if (!Matcher::isSimpleConstant64(lcon) && 4324 st->Opcode() == Op_StoreI) { 4325 // This StoreI is already optimal by itself. 4326 jint* intcon = (jint*) &tiles[j]; 4327 intcon[1] = 0; // undo the store_constant() 4328 4329 // If the previous store is also optimal by itself, back up and 4330 // undo the action of the previous loop iteration... if we can. 4331 // But if we can't, just let the previous half take care of itself. 4332 st = nodes[j]; 4333 st_off -= BytesPerInt; 4334 con = intcon[0]; 4335 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) { 4336 assert(st_off >= header_size, "still ignoring header"); 4337 assert(get_store_offset(st, phase) == st_off, "must be"); 4338 assert(in(i-1) == zmem, "must be"); 4339 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 4340 assert(con == tcon->is_int()->get_con(), "must be"); 4341 // Undo the effects of the previous loop trip, which swallowed st: 4342 intcon[0] = 0; // undo store_constant() 4343 set_req(i-1, st); // undo set_req(i, zmem) 4344 nodes[j] = nullptr; // undo nodes[j] = st 4345 --old_subword; // undo ++old_subword 4346 } 4347 continue; // This StoreI is already optimal. 4348 } 4349 } 4350 4351 // This store is not needed. 4352 set_req(i, zmem); 4353 nodes[j] = st; // record for the moment 4354 if (st_size < BytesPerLong) // something has changed 4355 ++old_subword; // includes int/float, but who's counting... 4356 else ++old_long; 4357 } 4358 4359 if ((old_subword + old_long) == 0) 4360 return; // nothing more to do 4361 4362 //// Pass B: Convert any non-zero tiles into optimal constant stores. 4363 // Be sure to insert them before overlapping non-constant stores. 4364 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 4365 for (int j = 0; j < num_tiles; j++) { 4366 jlong con = tiles[j]; 4367 jlong init = inits[j]; 4368 if (con == 0) continue; 4369 jint con0, con1; // split the constant, address-wise 4370 jint init0, init1; // split the init map, address-wise 4371 { union { jlong con; jint intcon[2]; } u; 4372 u.con = con; 4373 con0 = u.intcon[0]; 4374 con1 = u.intcon[1]; 4375 u.con = init; 4376 init0 = u.intcon[0]; 4377 init1 = u.intcon[1]; 4378 } 4379 4380 Node* old = nodes[j]; 4381 assert(old != nullptr, "need the prior store"); 4382 intptr_t offset = (j * BytesPerLong); 4383 4384 bool split = !Matcher::isSimpleConstant64(con); 4385 4386 if (offset < header_size) { 4387 assert(offset + BytesPerInt >= header_size, "second int counts"); 4388 assert(*(jint*)&tiles[j] == 0, "junk in header"); 4389 split = true; // only the second word counts 4390 // Example: int a[] = { 42 ... } 4391 } else if (con0 == 0 && init0 == -1) { 4392 split = true; // first word is covered by full inits 4393 // Example: int a[] = { ... foo(), 42 ... } 4394 } else if (con1 == 0 && init1 == -1) { 4395 split = true; // second word is covered by full inits 4396 // Example: int a[] = { ... 42, foo() ... } 4397 } 4398 4399 // Here's a case where init0 is neither 0 nor -1: 4400 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 4401 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 4402 // In this case the tile is not split; it is (jlong)42. 4403 // The big tile is stored down, and then the foo() value is inserted. 4404 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 4405 4406 Node* ctl = old->in(MemNode::Control); 4407 Node* adr = make_raw_address(offset, phase); 4408 const TypePtr* atp = TypeRawPtr::BOTTOM; 4409 4410 // One or two coalesced stores to plop down. 4411 Node* st[2]; 4412 intptr_t off[2]; 4413 int nst = 0; 4414 if (!split) { 4415 ++new_long; 4416 off[nst] = offset; 4417 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4418 phase->longcon(con), T_LONG, MemNode::unordered); 4419 } else { 4420 // Omit either if it is a zero. 4421 if (con0 != 0) { 4422 ++new_int; 4423 off[nst] = offset; 4424 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4425 phase->intcon(con0), T_INT, MemNode::unordered); 4426 } 4427 if (con1 != 0) { 4428 ++new_int; 4429 offset += BytesPerInt; 4430 adr = make_raw_address(offset, phase); 4431 off[nst] = offset; 4432 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4433 phase->intcon(con1), T_INT, MemNode::unordered); 4434 } 4435 } 4436 4437 // Insert second store first, then the first before the second. 4438 // Insert each one just before any overlapping non-constant stores. 4439 while (nst > 0) { 4440 Node* st1 = st[--nst]; 4441 C->copy_node_notes_to(st1, old); 4442 st1 = phase->transform(st1); 4443 offset = off[nst]; 4444 assert(offset >= header_size, "do not smash header"); 4445 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 4446 guarantee(ins_idx != 0, "must re-insert constant store"); 4447 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 4448 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 4449 set_req(--ins_idx, st1); 4450 else 4451 ins_req(ins_idx, st1); 4452 } 4453 } 4454 4455 if (PrintCompilation && WizardMode) 4456 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 4457 old_subword, old_long, new_int, new_long); 4458 if (C->log() != nullptr) 4459 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 4460 old_subword, old_long, new_int, new_long); 4461 4462 // Clean up any remaining occurrences of zmem: 4463 remove_extra_zeroes(); 4464 } 4465 4466 // Explore forward from in(start) to find the first fully initialized 4467 // word, and return its offset. Skip groups of subword stores which 4468 // together initialize full words. If in(start) is itself part of a 4469 // fully initialized word, return the offset of in(start). If there 4470 // are no following full-word stores, or if something is fishy, return 4471 // a negative value. 4472 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 4473 int int_map = 0; 4474 intptr_t int_map_off = 0; 4475 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 4476 4477 for (uint i = start, limit = req(); i < limit; i++) { 4478 Node* st = in(i); 4479 4480 intptr_t st_off = get_store_offset(st, phase); 4481 if (st_off < 0) break; // return conservative answer 4482 4483 int st_size = st->as_Store()->memory_size(); 4484 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 4485 return st_off; // we found a complete word init 4486 } 4487 4488 // update the map: 4489 4490 intptr_t this_int_off = align_down(st_off, BytesPerInt); 4491 if (this_int_off != int_map_off) { 4492 // reset the map: 4493 int_map = 0; 4494 int_map_off = this_int_off; 4495 } 4496 4497 int subword_off = st_off - this_int_off; 4498 int_map |= right_n_bits(st_size) << subword_off; 4499 if ((int_map & FULL_MAP) == FULL_MAP) { 4500 return this_int_off; // we found a complete word init 4501 } 4502 4503 // Did this store hit or cross the word boundary? 4504 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt); 4505 if (next_int_off == this_int_off + BytesPerInt) { 4506 // We passed the current int, without fully initializing it. 4507 int_map_off = next_int_off; 4508 int_map >>= BytesPerInt; 4509 } else if (next_int_off > this_int_off + BytesPerInt) { 4510 // We passed the current and next int. 4511 return this_int_off + BytesPerInt; 4512 } 4513 } 4514 4515 return -1; 4516 } 4517 4518 4519 // Called when the associated AllocateNode is expanded into CFG. 4520 // At this point, we may perform additional optimizations. 4521 // Linearize the stores by ascending offset, to make memory 4522 // activity as coherent as possible. 4523 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 4524 intptr_t header_size, 4525 Node* size_in_bytes, 4526 PhaseIterGVN* phase) { 4527 assert(!is_complete(), "not already complete"); 4528 assert(stores_are_sane(phase), ""); 4529 assert(allocation() != nullptr, "must be present"); 4530 4531 remove_extra_zeroes(); 4532 4533 if (ReduceFieldZeroing || ReduceBulkZeroing) 4534 // reduce instruction count for common initialization patterns 4535 coalesce_subword_stores(header_size, size_in_bytes, phase); 4536 4537 Node* zmem = zero_memory(); // initially zero memory state 4538 Node* inits = zmem; // accumulating a linearized chain of inits 4539 #ifdef ASSERT 4540 intptr_t first_offset = allocation()->minimum_header_size(); 4541 intptr_t last_init_off = first_offset; // previous init offset 4542 intptr_t last_init_end = first_offset; // previous init offset+size 4543 intptr_t last_tile_end = first_offset; // previous tile offset+size 4544 #endif 4545 intptr_t zeroes_done = header_size; 4546 4547 bool do_zeroing = true; // we might give up if inits are very sparse 4548 int big_init_gaps = 0; // how many large gaps have we seen? 4549 4550 if (UseTLAB && ZeroTLAB) do_zeroing = false; 4551 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 4552 4553 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4554 Node* st = in(i); 4555 intptr_t st_off = get_store_offset(st, phase); 4556 if (st_off < 0) 4557 break; // unknown junk in the inits 4558 if (st->in(MemNode::Memory) != zmem) 4559 break; // complicated store chains somehow in list 4560 4561 int st_size = st->as_Store()->memory_size(); 4562 intptr_t next_init_off = st_off + st_size; 4563 4564 if (do_zeroing && zeroes_done < next_init_off) { 4565 // See if this store needs a zero before it or under it. 4566 intptr_t zeroes_needed = st_off; 4567 4568 if (st_size < BytesPerInt) { 4569 // Look for subword stores which only partially initialize words. 4570 // If we find some, we must lay down some word-level zeroes first, 4571 // underneath the subword stores. 4572 // 4573 // Examples: 4574 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 4575 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 4576 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 4577 // 4578 // Note: coalesce_subword_stores may have already done this, 4579 // if it was prompted by constant non-zero subword initializers. 4580 // But this case can still arise with non-constant stores. 4581 4582 intptr_t next_full_store = find_next_fullword_store(i, phase); 4583 4584 // In the examples above: 4585 // in(i) p q r s x y z 4586 // st_off 12 13 14 15 12 13 14 4587 // st_size 1 1 1 1 1 1 1 4588 // next_full_s. 12 16 16 16 16 16 16 4589 // z's_done 12 16 16 16 12 16 12 4590 // z's_needed 12 16 16 16 16 16 16 4591 // zsize 0 0 0 0 4 0 4 4592 if (next_full_store < 0) { 4593 // Conservative tack: Zero to end of current word. 4594 zeroes_needed = align_up(zeroes_needed, BytesPerInt); 4595 } else { 4596 // Zero to beginning of next fully initialized word. 4597 // Or, don't zero at all, if we are already in that word. 4598 assert(next_full_store >= zeroes_needed, "must go forward"); 4599 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 4600 zeroes_needed = next_full_store; 4601 } 4602 } 4603 4604 if (zeroes_needed > zeroes_done) { 4605 intptr_t zsize = zeroes_needed - zeroes_done; 4606 // Do some incremental zeroing on rawmem, in parallel with inits. 4607 zeroes_done = align_down(zeroes_done, BytesPerInt); 4608 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4609 allocation()->in(AllocateNode::DefaultValue), 4610 allocation()->in(AllocateNode::RawDefaultValue), 4611 zeroes_done, zeroes_needed, 4612 phase); 4613 zeroes_done = zeroes_needed; 4614 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 4615 do_zeroing = false; // leave the hole, next time 4616 } 4617 } 4618 4619 // Collect the store and move on: 4620 phase->replace_input_of(st, MemNode::Memory, inits); 4621 inits = st; // put it on the linearized chain 4622 set_req(i, zmem); // unhook from previous position 4623 4624 if (zeroes_done == st_off) 4625 zeroes_done = next_init_off; 4626 4627 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 4628 4629 #ifdef ASSERT 4630 // Various order invariants. Weaker than stores_are_sane because 4631 // a large constant tile can be filled in by smaller non-constant stores. 4632 assert(st_off >= last_init_off, "inits do not reverse"); 4633 last_init_off = st_off; 4634 const Type* val = nullptr; 4635 if (st_size >= BytesPerInt && 4636 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 4637 (int)val->basic_type() < (int)T_OBJECT) { 4638 assert(st_off >= last_tile_end, "tiles do not overlap"); 4639 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 4640 last_tile_end = MAX2(last_tile_end, next_init_off); 4641 } else { 4642 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong); 4643 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 4644 assert(st_off >= last_init_end, "inits do not overlap"); 4645 last_init_end = next_init_off; // it's a non-tile 4646 } 4647 #endif //ASSERT 4648 } 4649 4650 remove_extra_zeroes(); // clear out all the zmems left over 4651 add_req(inits); 4652 4653 if (!(UseTLAB && ZeroTLAB)) { 4654 // If anything remains to be zeroed, zero it all now. 4655 zeroes_done = align_down(zeroes_done, BytesPerInt); 4656 // if it is the last unused 4 bytes of an instance, forget about it 4657 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 4658 if (zeroes_done + BytesPerLong >= size_limit) { 4659 AllocateNode* alloc = allocation(); 4660 assert(alloc != nullptr, "must be present"); 4661 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) { 4662 Node* klass_node = alloc->in(AllocateNode::KlassNode); 4663 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass(); 4664 if (zeroes_done == k->layout_helper()) 4665 zeroes_done = size_limit; 4666 } 4667 } 4668 if (zeroes_done < size_limit) { 4669 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4670 allocation()->in(AllocateNode::DefaultValue), 4671 allocation()->in(AllocateNode::RawDefaultValue), 4672 zeroes_done, size_in_bytes, phase); 4673 } 4674 } 4675 4676 set_complete(phase); 4677 return rawmem; 4678 } 4679 4680 4681 #ifdef ASSERT 4682 bool InitializeNode::stores_are_sane(PhaseValues* phase) { 4683 if (is_complete()) 4684 return true; // stores could be anything at this point 4685 assert(allocation() != nullptr, "must be present"); 4686 intptr_t last_off = allocation()->minimum_header_size(); 4687 for (uint i = InitializeNode::RawStores; i < req(); i++) { 4688 Node* st = in(i); 4689 intptr_t st_off = get_store_offset(st, phase); 4690 if (st_off < 0) continue; // ignore dead garbage 4691 if (last_off > st_off) { 4692 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 4693 this->dump(2); 4694 assert(false, "ascending store offsets"); 4695 return false; 4696 } 4697 last_off = st_off + st->as_Store()->memory_size(); 4698 } 4699 return true; 4700 } 4701 #endif //ASSERT 4702 4703 4704 4705 4706 //============================MergeMemNode===================================== 4707 // 4708 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4709 // contributing store or call operations. Each contributor provides the memory 4710 // state for a particular "alias type" (see Compile::alias_type). For example, 4711 // if a MergeMem has an input X for alias category #6, then any memory reference 4712 // to alias category #6 may use X as its memory state input, as an exact equivalent 4713 // to using the MergeMem as a whole. 4714 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4715 // 4716 // (Here, the <N> notation gives the index of the relevant adr_type.) 4717 // 4718 // In one special case (and more cases in the future), alias categories overlap. 4719 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4720 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4721 // it is exactly equivalent to that state W: 4722 // MergeMem(<Bot>: W) <==> W 4723 // 4724 // Usually, the merge has more than one input. In that case, where inputs 4725 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4726 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4727 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4728 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4729 // 4730 // A merge can take a "wide" memory state as one of its narrow inputs. 4731 // This simply means that the merge observes out only the relevant parts of 4732 // the wide input. That is, wide memory states arriving at narrow merge inputs 4733 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4734 // 4735 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4736 // and that memory slices "leak through": 4737 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4738 // 4739 // But, in such a cascade, repeated memory slices can "block the leak": 4740 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4741 // 4742 // In the last example, Y is not part of the combined memory state of the 4743 // outermost MergeMem. The system must, of course, prevent unschedulable 4744 // memory states from arising, so you can be sure that the state Y is somehow 4745 // a precursor to state Y'. 4746 // 4747 // 4748 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4749 // of each MergeMemNode array are exactly the numerical alias indexes, including 4750 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4751 // Compile::alias_type (and kin) produce and manage these indexes. 4752 // 4753 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4754 // (Note that this provides quick access to the top node inside MergeMem methods, 4755 // without the need to reach out via TLS to Compile::current.) 4756 // 4757 // As a consequence of what was just described, a MergeMem that represents a full 4758 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4759 // containing all alias categories. 4760 // 4761 // MergeMem nodes never (?) have control inputs, so in(0) is null. 4762 // 4763 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4764 // a memory state for the alias type <N>, or else the top node, meaning that 4765 // there is no particular input for that alias type. Note that the length of 4766 // a MergeMem is variable, and may be extended at any time to accommodate new 4767 // memory states at larger alias indexes. When merges grow, they are of course 4768 // filled with "top" in the unused in() positions. 4769 // 4770 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4771 // (Top was chosen because it works smoothly with passes like GCM.) 4772 // 4773 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4774 // the type of random VM bits like TLS references.) Since it is always the 4775 // first non-Bot memory slice, some low-level loops use it to initialize an 4776 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4777 // 4778 // 4779 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4780 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4781 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4782 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4783 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4784 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4785 // 4786 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4787 // really that different from the other memory inputs. An abbreviation called 4788 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4789 // 4790 // 4791 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4792 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4793 // that "emerges though" the base memory will be marked as excluding the alias types 4794 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4795 // 4796 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4797 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4798 // 4799 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4800 // (It is currently unimplemented.) As you can see, the resulting merge is 4801 // actually a disjoint union of memory states, rather than an overlay. 4802 // 4803 4804 //------------------------------MergeMemNode----------------------------------- 4805 Node* MergeMemNode::make_empty_memory() { 4806 Node* empty_memory = (Node*) Compile::current()->top(); 4807 assert(empty_memory->is_top(), "correct sentinel identity"); 4808 return empty_memory; 4809 } 4810 4811 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4812 init_class_id(Class_MergeMem); 4813 // all inputs are nullified in Node::Node(int) 4814 // set_input(0, nullptr); // no control input 4815 4816 // Initialize the edges uniformly to top, for starters. 4817 Node* empty_mem = make_empty_memory(); 4818 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4819 init_req(i,empty_mem); 4820 } 4821 assert(empty_memory() == empty_mem, ""); 4822 4823 if( new_base != nullptr && new_base->is_MergeMem() ) { 4824 MergeMemNode* mdef = new_base->as_MergeMem(); 4825 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4826 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4827 mms.set_memory(mms.memory2()); 4828 } 4829 assert(base_memory() == mdef->base_memory(), ""); 4830 } else { 4831 set_base_memory(new_base); 4832 } 4833 } 4834 4835 // Make a new, untransformed MergeMem with the same base as 'mem'. 4836 // If mem is itself a MergeMem, populate the result with the same edges. 4837 MergeMemNode* MergeMemNode::make(Node* mem) { 4838 return new MergeMemNode(mem); 4839 } 4840 4841 //------------------------------cmp-------------------------------------------- 4842 uint MergeMemNode::hash() const { return NO_HASH; } 4843 bool MergeMemNode::cmp( const Node &n ) const { 4844 return (&n == this); // Always fail except on self 4845 } 4846 4847 //------------------------------Identity--------------------------------------- 4848 Node* MergeMemNode::Identity(PhaseGVN* phase) { 4849 // Identity if this merge point does not record any interesting memory 4850 // disambiguations. 4851 Node* base_mem = base_memory(); 4852 Node* empty_mem = empty_memory(); 4853 if (base_mem != empty_mem) { // Memory path is not dead? 4854 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4855 Node* mem = in(i); 4856 if (mem != empty_mem && mem != base_mem) { 4857 return this; // Many memory splits; no change 4858 } 4859 } 4860 } 4861 return base_mem; // No memory splits; ID on the one true input 4862 } 4863 4864 //------------------------------Ideal------------------------------------------ 4865 // This method is invoked recursively on chains of MergeMem nodes 4866 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4867 // Remove chain'd MergeMems 4868 // 4869 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4870 // relative to the "in(Bot)". Since we are patching both at the same time, 4871 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4872 // but rewrite each "in(i)" relative to the new "in(Bot)". 4873 Node *progress = nullptr; 4874 4875 4876 Node* old_base = base_memory(); 4877 Node* empty_mem = empty_memory(); 4878 if (old_base == empty_mem) 4879 return nullptr; // Dead memory path. 4880 4881 MergeMemNode* old_mbase; 4882 if (old_base != nullptr && old_base->is_MergeMem()) 4883 old_mbase = old_base->as_MergeMem(); 4884 else 4885 old_mbase = nullptr; 4886 Node* new_base = old_base; 4887 4888 // simplify stacked MergeMems in base memory 4889 if (old_mbase) new_base = old_mbase->base_memory(); 4890 4891 // the base memory might contribute new slices beyond my req() 4892 if (old_mbase) grow_to_match(old_mbase); 4893 4894 // Note: We do not call verify_sparse on entry, because inputs 4895 // can normalize to the base_memory via subsume_node or similar 4896 // mechanisms. This method repairs that damage. 4897 4898 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4899 4900 // Look at each slice. 4901 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4902 Node* old_in = in(i); 4903 // calculate the old memory value 4904 Node* old_mem = old_in; 4905 if (old_mem == empty_mem) old_mem = old_base; 4906 assert(old_mem == memory_at(i), ""); 4907 4908 // maybe update (reslice) the old memory value 4909 4910 // simplify stacked MergeMems 4911 Node* new_mem = old_mem; 4912 MergeMemNode* old_mmem; 4913 if (old_mem != nullptr && old_mem->is_MergeMem()) 4914 old_mmem = old_mem->as_MergeMem(); 4915 else 4916 old_mmem = nullptr; 4917 if (old_mmem == this) { 4918 // This can happen if loops break up and safepoints disappear. 4919 // A merge of BotPtr (default) with a RawPtr memory derived from a 4920 // safepoint can be rewritten to a merge of the same BotPtr with 4921 // the BotPtr phi coming into the loop. If that phi disappears 4922 // also, we can end up with a self-loop of the mergemem. 4923 // In general, if loops degenerate and memory effects disappear, 4924 // a mergemem can be left looking at itself. This simply means 4925 // that the mergemem's default should be used, since there is 4926 // no longer any apparent effect on this slice. 4927 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4928 // from start. Update the input to TOP. 4929 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4930 } 4931 else if (old_mmem != nullptr) { 4932 new_mem = old_mmem->memory_at(i); 4933 } 4934 // else preceding memory was not a MergeMem 4935 4936 // maybe store down a new value 4937 Node* new_in = new_mem; 4938 if (new_in == new_base) new_in = empty_mem; 4939 4940 if (new_in != old_in) { 4941 // Warning: Do not combine this "if" with the previous "if" 4942 // A memory slice might have be be rewritten even if it is semantically 4943 // unchanged, if the base_memory value has changed. 4944 set_req_X(i, new_in, phase); 4945 progress = this; // Report progress 4946 } 4947 } 4948 4949 if (new_base != old_base) { 4950 set_req_X(Compile::AliasIdxBot, new_base, phase); 4951 // Don't use set_base_memory(new_base), because we need to update du. 4952 assert(base_memory() == new_base, ""); 4953 progress = this; 4954 } 4955 4956 if( base_memory() == this ) { 4957 // a self cycle indicates this memory path is dead 4958 set_req(Compile::AliasIdxBot, empty_mem); 4959 } 4960 4961 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4962 // Recursion must occur after the self cycle check above 4963 if( base_memory()->is_MergeMem() ) { 4964 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4965 Node *m = phase->transform(new_mbase); // Rollup any cycles 4966 if( m != nullptr && 4967 (m->is_top() || 4968 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) { 4969 // propagate rollup of dead cycle to self 4970 set_req(Compile::AliasIdxBot, empty_mem); 4971 } 4972 } 4973 4974 if( base_memory() == empty_mem ) { 4975 progress = this; 4976 // Cut inputs during Parse phase only. 4977 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4978 if( !can_reshape ) { 4979 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4980 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4981 } 4982 } 4983 } 4984 4985 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4986 // Check if PhiNode::Ideal's "Split phis through memory merges" 4987 // transform should be attempted. Look for this->phi->this cycle. 4988 uint merge_width = req(); 4989 if (merge_width > Compile::AliasIdxRaw) { 4990 PhiNode* phi = base_memory()->as_Phi(); 4991 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4992 if (phi->in(i) == this) { 4993 phase->is_IterGVN()->_worklist.push(phi); 4994 break; 4995 } 4996 } 4997 } 4998 } 4999 5000 assert(progress || verify_sparse(), "please, no dups of base"); 5001 return progress; 5002 } 5003 5004 //-------------------------set_base_memory------------------------------------- 5005 void MergeMemNode::set_base_memory(Node *new_base) { 5006 Node* empty_mem = empty_memory(); 5007 set_req(Compile::AliasIdxBot, new_base); 5008 assert(memory_at(req()) == new_base, "must set default memory"); 5009 // Clear out other occurrences of new_base: 5010 if (new_base != empty_mem) { 5011 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 5012 if (in(i) == new_base) set_req(i, empty_mem); 5013 } 5014 } 5015 } 5016 5017 //------------------------------out_RegMask------------------------------------ 5018 const RegMask &MergeMemNode::out_RegMask() const { 5019 return RegMask::Empty; 5020 } 5021 5022 //------------------------------dump_spec-------------------------------------- 5023 #ifndef PRODUCT 5024 void MergeMemNode::dump_spec(outputStream *st) const { 5025 st->print(" {"); 5026 Node* base_mem = base_memory(); 5027 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 5028 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem; 5029 if (mem == base_mem) { st->print(" -"); continue; } 5030 st->print( " N%d:", mem->_idx ); 5031 Compile::current()->get_adr_type(i)->dump_on(st); 5032 } 5033 st->print(" }"); 5034 } 5035 #endif // !PRODUCT 5036 5037 5038 #ifdef ASSERT 5039 static bool might_be_same(Node* a, Node* b) { 5040 if (a == b) return true; 5041 if (!(a->is_Phi() || b->is_Phi())) return false; 5042 // phis shift around during optimization 5043 return true; // pretty stupid... 5044 } 5045 5046 // verify a narrow slice (either incoming or outgoing) 5047 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 5048 if (!VerifyAliases) return; // don't bother to verify unless requested 5049 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error 5050 if (Node::in_dump()) return; // muzzle asserts when printing 5051 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 5052 assert(n != nullptr, ""); 5053 // Elide intervening MergeMem's 5054 while (n->is_MergeMem()) { 5055 n = n->as_MergeMem()->memory_at(alias_idx); 5056 } 5057 Compile* C = Compile::current(); 5058 const TypePtr* n_adr_type = n->adr_type(); 5059 if (n == m->empty_memory()) { 5060 // Implicit copy of base_memory() 5061 } else if (n_adr_type != TypePtr::BOTTOM) { 5062 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type"); 5063 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 5064 } else { 5065 // A few places like make_runtime_call "know" that VM calls are narrow, 5066 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 5067 bool expected_wide_mem = false; 5068 if (n == m->base_memory()) { 5069 expected_wide_mem = true; 5070 } else if (alias_idx == Compile::AliasIdxRaw || 5071 n == m->memory_at(Compile::AliasIdxRaw)) { 5072 expected_wide_mem = true; 5073 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 5074 // memory can "leak through" calls on channels that 5075 // are write-once. Allow this also. 5076 expected_wide_mem = true; 5077 } 5078 assert(expected_wide_mem, "expected narrow slice replacement"); 5079 } 5080 } 5081 #else // !ASSERT 5082 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 5083 #endif 5084 5085 5086 //-----------------------------memory_at--------------------------------------- 5087 Node* MergeMemNode::memory_at(uint alias_idx) const { 5088 assert(alias_idx >= Compile::AliasIdxRaw || 5089 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()), 5090 "must avoid base_memory and AliasIdxTop"); 5091 5092 // Otherwise, it is a narrow slice. 5093 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 5094 if (is_empty_memory(n)) { 5095 // the array is sparse; empty slots are the "top" node 5096 n = base_memory(); 5097 assert(Node::in_dump() 5098 || n == nullptr || n->bottom_type() == Type::TOP 5099 || n->adr_type() == nullptr // address is TOP 5100 || n->adr_type() == TypePtr::BOTTOM 5101 || n->adr_type() == TypeRawPtr::BOTTOM 5102 || !Compile::current()->do_aliasing(), 5103 "must be a wide memory"); 5104 // do_aliasing == false if we are organizing the memory states manually. 5105 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 5106 } else { 5107 // make sure the stored slice is sane 5108 #ifdef ASSERT 5109 if (VMError::is_error_reported() || Node::in_dump()) { 5110 } else if (might_be_same(n, base_memory())) { 5111 // Give it a pass: It is a mostly harmless repetition of the base. 5112 // This can arise normally from node subsumption during optimization. 5113 } else { 5114 verify_memory_slice(this, alias_idx, n); 5115 } 5116 #endif 5117 } 5118 return n; 5119 } 5120 5121 //---------------------------set_memory_at------------------------------------- 5122 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 5123 verify_memory_slice(this, alias_idx, n); 5124 Node* empty_mem = empty_memory(); 5125 if (n == base_memory()) n = empty_mem; // collapse default 5126 uint need_req = alias_idx+1; 5127 if (req() < need_req) { 5128 if (n == empty_mem) return; // already the default, so do not grow me 5129 // grow the sparse array 5130 do { 5131 add_req(empty_mem); 5132 } while (req() < need_req); 5133 } 5134 set_req( alias_idx, n ); 5135 } 5136 5137 5138 5139 //--------------------------iteration_setup------------------------------------ 5140 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 5141 if (other != nullptr) { 5142 grow_to_match(other); 5143 // invariant: the finite support of mm2 is within mm->req() 5144 #ifdef ASSERT 5145 for (uint i = req(); i < other->req(); i++) { 5146 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 5147 } 5148 #endif 5149 } 5150 // Replace spurious copies of base_memory by top. 5151 Node* base_mem = base_memory(); 5152 if (base_mem != nullptr && !base_mem->is_top()) { 5153 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 5154 if (in(i) == base_mem) 5155 set_req(i, empty_memory()); 5156 } 5157 } 5158 } 5159 5160 //---------------------------grow_to_match------------------------------------- 5161 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 5162 Node* empty_mem = empty_memory(); 5163 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 5164 // look for the finite support of the other memory 5165 for (uint i = other->req(); --i >= req(); ) { 5166 if (other->in(i) != empty_mem) { 5167 uint new_len = i+1; 5168 while (req() < new_len) add_req(empty_mem); 5169 break; 5170 } 5171 } 5172 } 5173 5174 //---------------------------verify_sparse------------------------------------- 5175 #ifndef PRODUCT 5176 bool MergeMemNode::verify_sparse() const { 5177 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 5178 Node* base_mem = base_memory(); 5179 // The following can happen in degenerate cases, since empty==top. 5180 if (is_empty_memory(base_mem)) return true; 5181 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 5182 assert(in(i) != nullptr, "sane slice"); 5183 if (in(i) == base_mem) return false; // should have been the sentinel value! 5184 } 5185 return true; 5186 } 5187 5188 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 5189 Node* n; 5190 n = mm->in(idx); 5191 if (mem == n) return true; // might be empty_memory() 5192 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 5193 if (mem == n) return true; 5194 return false; 5195 } 5196 #endif // !PRODUCT