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