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