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