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