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