1 /* 2 * Copyright (c) 1997, 2024, 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->isa_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(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(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)); 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); 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), 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 uint8_t MemNode::barrier_data(const Node* n) { 843 if (n->is_LoadStore()) { 844 return n->as_LoadStore()->barrier_data(); 845 } else if (n->is_Mem()) { 846 return n->as_Mem()->barrier_data(); 847 } 848 return 0; 849 } 850 851 //============================================================================= 852 // Should LoadNode::Ideal() attempt to remove control edges? 853 bool LoadNode::can_remove_control() const { 854 return !has_pinned_control_dependency(); 855 } 856 uint LoadNode::size_of() const { return sizeof(*this); } 857 bool LoadNode::cmp(const Node &n) const { 858 LoadNode& load = (LoadNode &)n; 859 return !Type::cmp(_type, load._type) && 860 _control_dependency == load._control_dependency && 861 _mo == load._mo; 862 } 863 const Type *LoadNode::bottom_type() const { return _type; } 864 uint LoadNode::ideal_reg() const { 865 return _type->ideal_reg(); 866 } 867 868 #ifndef PRODUCT 869 void LoadNode::dump_spec(outputStream *st) const { 870 MemNode::dump_spec(st); 871 if( !Verbose && !WizardMode ) { 872 // standard dump does this in Verbose and WizardMode 873 st->print(" #"); _type->dump_on(st); 874 } 875 if (!depends_only_on_test()) { 876 st->print(" (does not depend only on test, "); 877 if (control_dependency() == UnknownControl) { 878 st->print("unknown control"); 879 } else if (control_dependency() == Pinned) { 880 st->print("pinned"); 881 } else if (adr_type() == TypeRawPtr::BOTTOM) { 882 st->print("raw access"); 883 } else { 884 st->print("unknown reason"); 885 } 886 st->print(")"); 887 } 888 } 889 #endif 890 891 #ifdef ASSERT 892 //----------------------------is_immutable_value------------------------------- 893 // Helper function to allow a raw load without control edge for some cases 894 bool LoadNode::is_immutable_value(Node* adr) { 895 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 896 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) { 897 898 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1); 899 int offsets[] = { 900 in_bytes(JavaThread::osthread_offset()), 901 in_bytes(JavaThread::threadObj_offset()), 902 in_bytes(JavaThread::vthread_offset()), 903 in_bytes(JavaThread::scopedValueCache_offset()), 904 }; 905 906 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) { 907 if (offset == offsets[i]) { 908 return true; 909 } 910 } 911 } 912 913 return false; 914 } 915 #endif 916 917 //----------------------------LoadNode::make----------------------------------- 918 // Polymorphic factory method: 919 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo, 920 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) { 921 Compile* C = gvn.C; 922 923 // sanity check the alias category against the created node type 924 assert(!(adr_type->isa_oopptr() && 925 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 926 "use LoadKlassNode instead"); 927 assert(!(adr_type->isa_aryptr() && 928 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 929 "use LoadRangeNode instead"); 930 // Check control edge of raw loads 931 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 932 // oop will be recorded in oop map if load crosses safepoint 933 rt->isa_oopptr() || is_immutable_value(adr), 934 "raw memory operations should have control edge"); 935 LoadNode* load = nullptr; 936 switch (bt) { 937 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 938 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 939 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 940 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 941 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break; 942 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break; 943 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break; 944 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break; 945 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break; 946 case T_OBJECT: 947 case T_NARROWOOP: 948 #ifdef _LP64 949 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 950 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency); 951 } else 952 #endif 953 { 954 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 955 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); 956 } 957 break; 958 default: 959 ShouldNotReachHere(); 960 break; 961 } 962 assert(load != nullptr, "LoadNode should have been created"); 963 if (unaligned) { 964 load->set_unaligned_access(); 965 } 966 if (mismatched) { 967 load->set_mismatched_access(); 968 } 969 if (unsafe) { 970 load->set_unsafe_access(); 971 } 972 load->set_barrier_data(barrier_data); 973 if (load->Opcode() == Op_LoadN) { 974 Node* ld = gvn.transform(load); 975 return new DecodeNNode(ld, ld->bottom_type()->make_ptr()); 976 } 977 978 return load; 979 } 980 981 //------------------------------hash------------------------------------------- 982 uint LoadNode::hash() const { 983 // unroll addition of interesting fields 984 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 985 } 986 987 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 988 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) { 989 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile(); 990 bool is_stable_ary = FoldStableValues && 991 (tp != nullptr) && (tp->isa_aryptr() != nullptr) && 992 tp->isa_aryptr()->is_stable(); 993 994 return (eliminate_boxing && non_volatile) || is_stable_ary; 995 } 996 997 return false; 998 } 999 1000 LoadNode* LoadNode::pin_array_access_node() const { 1001 const TypePtr* adr_type = this->adr_type(); 1002 if (adr_type != nullptr && adr_type->isa_aryptr()) { 1003 return clone_pinned(); 1004 } 1005 return nullptr; 1006 } 1007 1008 // Is the value loaded previously stored by an arraycopy? If so return 1009 // a load node that reads from the source array so we may be able to 1010 // optimize out the ArrayCopy node later. 1011 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const { 1012 Node* ld_adr = in(MemNode::Address); 1013 intptr_t ld_off = 0; 1014 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 1015 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true); 1016 if (ac != nullptr) { 1017 assert(ac->is_ArrayCopy(), "what kind of node can this be?"); 1018 1019 Node* mem = ac->in(TypeFunc::Memory); 1020 Node* ctl = ac->in(0); 1021 Node* src = ac->in(ArrayCopyNode::Src); 1022 1023 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) { 1024 return nullptr; 1025 } 1026 1027 // load depends on the tests that validate the arraycopy 1028 LoadNode* ld = clone_pinned(); 1029 Node* addp = in(MemNode::Address)->clone(); 1030 if (ac->as_ArrayCopy()->is_clonebasic()) { 1031 assert(ld_alloc != nullptr, "need an alloc"); 1032 assert(addp->is_AddP(), "address must be addp"); 1033 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1034 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern"); 1035 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern"); 1036 addp->set_req(AddPNode::Base, src); 1037 addp->set_req(AddPNode::Address, src); 1038 } else { 1039 assert(ac->as_ArrayCopy()->is_arraycopy_validated() || 1040 ac->as_ArrayCopy()->is_copyof_validated() || 1041 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases"); 1042 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be"); 1043 addp->set_req(AddPNode::Base, src); 1044 addp->set_req(AddPNode::Address, src); 1045 1046 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr(); 1047 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type(); 1048 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT; 1049 1050 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem); 1051 uint shift = exact_log2(type2aelembytes(ary_elem)); 1052 1053 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 1054 #ifdef _LP64 1055 diff = phase->transform(new ConvI2LNode(diff)); 1056 #endif 1057 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift))); 1058 1059 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff)); 1060 addp->set_req(AddPNode::Offset, offset); 1061 } 1062 addp = phase->transform(addp); 1063 #ifdef ASSERT 1064 const TypePtr* adr_type = phase->type(addp)->is_ptr(); 1065 ld->_adr_type = adr_type; 1066 #endif 1067 ld->set_req(MemNode::Address, addp); 1068 ld->set_req(0, ctl); 1069 ld->set_req(MemNode::Memory, mem); 1070 return ld; 1071 } 1072 return nullptr; 1073 } 1074 1075 1076 //---------------------------can_see_stored_value------------------------------ 1077 // This routine exists to make sure this set of tests is done the same 1078 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 1079 // will change the graph shape in a way which makes memory alive twice at the 1080 // same time (uses the Oracle model of aliasing), then some 1081 // LoadXNode::Identity will fold things back to the equivalence-class model 1082 // of aliasing. 1083 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const { 1084 Node* ld_adr = in(MemNode::Address); 1085 intptr_t ld_off = 0; 1086 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off); 1087 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base); 1088 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 1089 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr; 1090 // This is more general than load from boxing objects. 1091 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 1092 uint alias_idx = atp->index(); 1093 Node* result = nullptr; 1094 Node* current = st; 1095 // Skip through chains of MemBarNodes checking the MergeMems for 1096 // new states for the slice of this load. Stop once any other 1097 // kind of node is encountered. Loads from final memory can skip 1098 // through any kind of MemBar but normal loads shouldn't skip 1099 // through MemBarAcquire since the could allow them to move out of 1100 // a synchronized region. It is not safe to step over MemBarCPUOrder, 1101 // because alias info above them may be inaccurate (e.g., due to 1102 // mixed/mismatched unsafe accesses). 1103 bool is_final_mem = !atp->is_rewritable(); 1104 while (current->is_Proj()) { 1105 int opc = current->in(0)->Opcode(); 1106 if ((is_final_mem && (opc == Op_MemBarAcquire || 1107 opc == Op_MemBarAcquireLock || 1108 opc == Op_LoadFence)) || 1109 opc == Op_MemBarRelease || 1110 opc == Op_StoreFence || 1111 opc == Op_MemBarReleaseLock || 1112 opc == Op_MemBarStoreStore || 1113 opc == Op_StoreStoreFence) { 1114 Node* mem = current->in(0)->in(TypeFunc::Memory); 1115 if (mem->is_MergeMem()) { 1116 MergeMemNode* merge = mem->as_MergeMem(); 1117 Node* new_st = merge->memory_at(alias_idx); 1118 if (new_st == merge->base_memory()) { 1119 // Keep searching 1120 current = new_st; 1121 continue; 1122 } 1123 // Save the new memory state for the slice and fall through 1124 // to exit. 1125 result = new_st; 1126 } 1127 } 1128 break; 1129 } 1130 if (result != nullptr) { 1131 st = result; 1132 } 1133 } 1134 1135 // Loop around twice in the case Load -> Initialize -> Store. 1136 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 1137 for (int trip = 0; trip <= 1; trip++) { 1138 1139 if (st->is_Store()) { 1140 Node* st_adr = st->in(MemNode::Address); 1141 if (st_adr != ld_adr) { 1142 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers). 1143 intptr_t st_off = 0; 1144 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off); 1145 if (ld_base == nullptr) return nullptr; 1146 if (st_base == nullptr) return nullptr; 1147 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr; 1148 if (ld_off != st_off) return nullptr; 1149 if (ld_off == Type::OffsetBot) return nullptr; 1150 // Same base, same offset. 1151 // Possible improvement for arrays: check index value instead of absolute offset. 1152 1153 // At this point we have proven something like this setup: 1154 // B = << base >> 1155 // L = LoadQ(AddP(Check/CastPP(B), #Off)) 1156 // S = StoreQ(AddP( B , #Off), V) 1157 // (Actually, we haven't yet proven the Q's are the same.) 1158 // In other words, we are loading from a casted version of 1159 // the same pointer-and-offset that we stored to. 1160 // Casted version may carry a dependency and it is respected. 1161 // Thus, we are able to replace L by V. 1162 } 1163 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1164 if (store_Opcode() != st->Opcode()) { 1165 return nullptr; 1166 } 1167 // LoadVector/StoreVector needs additional check to ensure the types match. 1168 if (st->is_StoreVector()) { 1169 const TypeVect* in_vt = st->as_StoreVector()->vect_type(); 1170 const TypeVect* out_vt = as_LoadVector()->vect_type(); 1171 if (in_vt != out_vt) { 1172 return nullptr; 1173 } 1174 } 1175 return st->in(MemNode::ValueIn); 1176 } 1177 1178 // A load from a freshly-created object always returns zero. 1179 // (This can happen after LoadNode::Ideal resets the load's memory input 1180 // to find_captured_store, which returned InitializeNode::zero_memory.) 1181 if (st->is_Proj() && st->in(0)->is_Allocate() && 1182 (st->in(0) == ld_alloc) && 1183 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1184 // return a zero value for the load's basic type 1185 // (This is one of the few places where a generic PhaseTransform 1186 // can create new nodes. Think of it as lazily manifesting 1187 // virtually pre-existing constants.) 1188 if (memory_type() != T_VOID) { 1189 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) { 1190 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an 1191 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done 1192 // by the ArrayCopyNode. 1193 return phase->zerocon(memory_type()); 1194 } 1195 } else { 1196 // TODO: materialize all-zero vector constant 1197 assert(!isa_Load() || as_Load()->type()->isa_vect(), ""); 1198 } 1199 } 1200 1201 // A load from an initialization barrier can match a captured store. 1202 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1203 InitializeNode* init = st->in(0)->as_Initialize(); 1204 AllocateNode* alloc = init->allocation(); 1205 if ((alloc != nullptr) && (alloc == ld_alloc)) { 1206 // examine a captured store value 1207 st = init->find_captured_store(ld_off, memory_size(), phase); 1208 if (st != nullptr) { 1209 continue; // take one more trip around 1210 } 1211 } 1212 } 1213 1214 // Load boxed value from result of valueOf() call is input parameter. 1215 if (this->is_Load() && ld_adr->is_AddP() && 1216 (tp != nullptr) && tp->is_ptr_to_boxed_value()) { 1217 intptr_t ignore = 0; 1218 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1219 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1220 base = bs->step_over_gc_barrier(base); 1221 if (base != nullptr && base->is_Proj() && 1222 base->as_Proj()->_con == TypeFunc::Parms && 1223 base->in(0)->is_CallStaticJava() && 1224 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1225 return base->in(0)->in(TypeFunc::Parms); 1226 } 1227 } 1228 1229 break; 1230 } 1231 1232 return nullptr; 1233 } 1234 1235 //----------------------is_instance_field_load_with_local_phi------------------ 1236 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1237 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1238 in(Address)->is_AddP() ) { 1239 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1240 // Only instances and boxed values. 1241 if( t_oop != nullptr && 1242 (t_oop->is_ptr_to_boxed_value() || 1243 t_oop->is_known_instance_field()) && 1244 t_oop->offset() != Type::OffsetBot && 1245 t_oop->offset() != Type::OffsetTop) { 1246 return true; 1247 } 1248 } 1249 return false; 1250 } 1251 1252 //------------------------------Identity--------------------------------------- 1253 // Loads are identity if previous store is to same address 1254 Node* LoadNode::Identity(PhaseGVN* phase) { 1255 // If the previous store-maker is the right kind of Store, and the store is 1256 // to the same address, then we are equal to the value stored. 1257 Node* mem = in(Memory); 1258 Node* value = can_see_stored_value(mem, phase); 1259 if( value ) { 1260 // byte, short & char stores truncate naturally. 1261 // A load has to load the truncated value which requires 1262 // some sort of masking operation and that requires an 1263 // Ideal call instead of an Identity call. 1264 if (memory_size() < BytesPerInt) { 1265 // If the input to the store does not fit with the load's result type, 1266 // it must be truncated via an Ideal call. 1267 if (!phase->type(value)->higher_equal(phase->type(this))) 1268 return this; 1269 } 1270 // (This works even when value is a Con, but LoadNode::Value 1271 // usually runs first, producing the singleton type of the Con.) 1272 if (!has_pinned_control_dependency() || value->is_Con()) { 1273 return value; 1274 } else { 1275 return this; 1276 } 1277 } 1278 1279 if (has_pinned_control_dependency()) { 1280 return this; 1281 } 1282 // Search for an existing data phi which was generated before for the same 1283 // instance's field to avoid infinite generation of phis in a loop. 1284 Node *region = mem->in(0); 1285 if (is_instance_field_load_with_local_phi(region)) { 1286 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1287 int this_index = phase->C->get_alias_index(addr_t); 1288 int this_offset = addr_t->offset(); 1289 int this_iid = addr_t->instance_id(); 1290 if (!addr_t->is_known_instance() && 1291 addr_t->is_ptr_to_boxed_value()) { 1292 // Use _idx of address base (could be Phi node) for boxed values. 1293 intptr_t ignore = 0; 1294 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1295 if (base == nullptr) { 1296 return this; 1297 } 1298 this_iid = base->_idx; 1299 } 1300 const Type* this_type = bottom_type(); 1301 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1302 Node* phi = region->fast_out(i); 1303 if (phi->is_Phi() && phi != mem && 1304 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) { 1305 return phi; 1306 } 1307 } 1308 } 1309 1310 return this; 1311 } 1312 1313 // Construct an equivalent unsigned load. 1314 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) { 1315 BasicType bt = T_ILLEGAL; 1316 const Type* rt = nullptr; 1317 switch (Opcode()) { 1318 case Op_LoadUB: return this; 1319 case Op_LoadUS: return this; 1320 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break; 1321 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break; 1322 default: 1323 assert(false, "no unsigned variant: %s", Name()); 1324 return nullptr; 1325 } 1326 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1327 raw_adr_type(), rt, bt, _mo, _control_dependency, 1328 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access()); 1329 } 1330 1331 // Construct an equivalent signed load. 1332 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) { 1333 BasicType bt = T_ILLEGAL; 1334 const Type* rt = nullptr; 1335 switch (Opcode()) { 1336 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break; 1337 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break; 1338 case Op_LoadB: // fall through 1339 case Op_LoadS: // fall through 1340 case Op_LoadI: // fall through 1341 case Op_LoadL: return this; 1342 default: 1343 assert(false, "no signed variant: %s", Name()); 1344 return nullptr; 1345 } 1346 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1347 raw_adr_type(), rt, bt, _mo, _control_dependency, 1348 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access()); 1349 } 1350 1351 bool LoadNode::has_reinterpret_variant(const Type* rt) { 1352 BasicType bt = rt->basic_type(); 1353 switch (Opcode()) { 1354 case Op_LoadI: return (bt == T_FLOAT); 1355 case Op_LoadL: return (bt == T_DOUBLE); 1356 case Op_LoadF: return (bt == T_INT); 1357 case Op_LoadD: return (bt == T_LONG); 1358 1359 default: return false; 1360 } 1361 } 1362 1363 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) { 1364 BasicType bt = rt->basic_type(); 1365 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt)); 1366 bool is_mismatched = is_mismatched_access(); 1367 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr(); 1368 if (raw_type == nullptr) { 1369 is_mismatched = true; // conservatively match all non-raw accesses as mismatched 1370 } 1371 const int op = Opcode(); 1372 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) || 1373 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access()); 1374 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1375 raw_adr_type(), rt, bt, _mo, _control_dependency, 1376 require_atomic_access, is_unaligned_access(), is_mismatched); 1377 } 1378 1379 bool StoreNode::has_reinterpret_variant(const Type* vt) { 1380 BasicType bt = vt->basic_type(); 1381 switch (Opcode()) { 1382 case Op_StoreI: return (bt == T_FLOAT); 1383 case Op_StoreL: return (bt == T_DOUBLE); 1384 case Op_StoreF: return (bt == T_INT); 1385 case Op_StoreD: return (bt == T_LONG); 1386 1387 default: return false; 1388 } 1389 } 1390 1391 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) { 1392 BasicType bt = vt->basic_type(); 1393 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt)); 1394 const int op = Opcode(); 1395 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) || 1396 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access()); 1397 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), 1398 raw_adr_type(), val, bt, _mo, require_atomic_access); 1399 1400 bool is_mismatched = is_mismatched_access(); 1401 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr(); 1402 if (raw_type == nullptr) { 1403 is_mismatched = true; // conservatively match all non-raw accesses as mismatched 1404 } 1405 if (is_mismatched) { 1406 st->set_mismatched_access(); 1407 } 1408 return st; 1409 } 1410 1411 // We're loading from an object which has autobox behaviour. 1412 // If this object is result of a valueOf call we'll have a phi 1413 // merging a newly allocated object and a load from the cache. 1414 // We want to replace this load with the original incoming 1415 // argument to the valueOf call. 1416 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) { 1417 assert(igvn->C->eliminate_boxing(), "sanity"); 1418 intptr_t ignore = 0; 1419 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore); 1420 if ((base == nullptr) || base->is_Phi()) { 1421 // Push the loads from the phi that comes from valueOf up 1422 // through it to allow elimination of the loads and the recovery 1423 // of the original value. It is done in split_through_phi(). 1424 return nullptr; 1425 } else if (base->is_Load() || 1426 (base->is_DecodeN() && base->in(1)->is_Load())) { 1427 // Eliminate the load of boxed value for integer types from the cache 1428 // array by deriving the value from the index into the array. 1429 // Capture the offset of the load and then reverse the computation. 1430 1431 // Get LoadN node which loads a boxing object from 'cache' array. 1432 if (base->is_DecodeN()) { 1433 base = base->in(1); 1434 } 1435 if (!base->in(Address)->is_AddP()) { 1436 return nullptr; // Complex address 1437 } 1438 AddPNode* address = base->in(Address)->as_AddP(); 1439 Node* cache_base = address->in(AddPNode::Base); 1440 if ((cache_base != nullptr) && cache_base->is_DecodeN()) { 1441 // Get ConP node which is static 'cache' field. 1442 cache_base = cache_base->in(1); 1443 } 1444 if ((cache_base != nullptr) && cache_base->is_Con()) { 1445 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1446 if ((base_type != nullptr) && base_type->is_autobox_cache()) { 1447 Node* elements[4]; 1448 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1449 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1450 if (count > 0 && elements[0]->is_Con() && 1451 (count == 1 || 1452 (count == 2 && elements[1]->Opcode() == Op_LShiftX && 1453 elements[1]->in(2) == igvn->intcon(shift)))) { 1454 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1455 // Fetch the box object cache[0] at the base of the array and get its value 1456 ciInstance* box = array->obj_at(0)->as_instance(); 1457 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1458 assert(ik->is_box_klass(), "sanity"); 1459 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1460 if (ik->nof_nonstatic_fields() == 1) { 1461 // This should be true nonstatic_field_at requires calling 1462 // nof_nonstatic_fields so check it anyway 1463 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1464 BasicType bt = c.basic_type(); 1465 // Only integer types have boxing cache. 1466 assert(bt == T_BOOLEAN || bt == T_CHAR || 1467 bt == T_BYTE || bt == T_SHORT || 1468 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt)); 1469 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1470 if (cache_low != (int)cache_low) { 1471 return nullptr; // should not happen since cache is array indexed by value 1472 } 1473 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1474 if (offset != (int)offset) { 1475 return nullptr; // should not happen since cache is array indexed by value 1476 } 1477 // Add up all the offsets making of the address of the load 1478 Node* result = elements[0]; 1479 for (int i = 1; i < count; i++) { 1480 result = igvn->transform(new AddXNode(result, elements[i])); 1481 } 1482 // Remove the constant offset from the address and then 1483 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset))); 1484 // remove the scaling of the offset to recover the original index. 1485 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) { 1486 // Peel the shift off directly but wrap it in a dummy node 1487 // since Ideal can't return existing nodes 1488 igvn->_worklist.push(result); // remove dead node later 1489 result = new RShiftXNode(result->in(1), igvn->intcon(0)); 1490 } else if (result->is_Add() && result->in(2)->is_Con() && 1491 result->in(1)->Opcode() == Op_LShiftX && 1492 result->in(1)->in(2) == igvn->intcon(shift)) { 1493 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1494 // but for boxing cache access we know that X<<Z will not overflow 1495 // (there is range check) so we do this optimizatrion by hand here. 1496 igvn->_worklist.push(result); // remove dead node later 1497 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift)); 1498 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con)); 1499 } else { 1500 result = new RShiftXNode(result, igvn->intcon(shift)); 1501 } 1502 #ifdef _LP64 1503 if (bt != T_LONG) { 1504 result = new ConvL2INode(igvn->transform(result)); 1505 } 1506 #else 1507 if (bt == T_LONG) { 1508 result = new ConvI2LNode(igvn->transform(result)); 1509 } 1510 #endif 1511 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair). 1512 // Need to preserve unboxing load type if it is unsigned. 1513 switch(this->Opcode()) { 1514 case Op_LoadUB: 1515 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF)); 1516 break; 1517 case Op_LoadUS: 1518 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF)); 1519 break; 1520 } 1521 return result; 1522 } 1523 } 1524 } 1525 } 1526 } 1527 return nullptr; 1528 } 1529 1530 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1531 Node* region = phi->in(0); 1532 if (region == nullptr) { 1533 return false; // Wait stable graph 1534 } 1535 uint cnt = phi->req(); 1536 for (uint i = 1; i < cnt; i++) { 1537 Node* rc = region->in(i); 1538 if (rc == nullptr || phase->type(rc) == Type::TOP) 1539 return false; // Wait stable graph 1540 Node* in = phi->in(i); 1541 if (in == nullptr || phase->type(in) == Type::TOP) 1542 return false; // Wait stable graph 1543 } 1544 return true; 1545 } 1546 1547 //------------------------------split_through_phi------------------------------ 1548 // Check whether a call to 'split_through_phi' would split this load through the 1549 // Phi *base*. This method is essentially a copy of the validations performed 1550 // by 'split_through_phi'. The first use of this method was in EA code as part 1551 // of simplification of allocation merges. 1552 // Some differences from original method (split_through_phi): 1553 // - If base->is_CastPP(): base = base->in(1) 1554 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) { 1555 Node* mem = in(Memory); 1556 Node* address = in(Address); 1557 intptr_t ignore = 0; 1558 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1559 1560 if (base->is_CastPP()) { 1561 base = base->in(1); 1562 } 1563 1564 if (req() > 3 || base == nullptr || !base->is_Phi()) { 1565 return false; 1566 } 1567 1568 if (!mem->is_Phi()) { 1569 if (!MemNode::all_controls_dominate(mem, base->in(0))) 1570 return false; 1571 } else if (base->in(0) != mem->in(0)) { 1572 if (!MemNode::all_controls_dominate(mem, base->in(0))) { 1573 return false; 1574 } 1575 } 1576 1577 return true; 1578 } 1579 1580 //------------------------------split_through_phi------------------------------ 1581 // Split instance or boxed field load through Phi. 1582 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) { 1583 if (req() > 3) { 1584 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs"); 1585 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account 1586 return nullptr; 1587 } 1588 Node* mem = in(Memory); 1589 Node* address = in(Address); 1590 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1591 1592 assert((t_oop != nullptr) && 1593 (ignore_missing_instance_id || 1594 t_oop->is_known_instance_field() || 1595 t_oop->is_ptr_to_boxed_value()), "invalid conditions"); 1596 1597 Compile* C = phase->C; 1598 intptr_t ignore = 0; 1599 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1600 bool base_is_phi = (base != nullptr) && base->is_Phi(); 1601 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1602 (base != nullptr) && (base == address->in(AddPNode::Base)) && 1603 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1604 1605 if (!((mem->is_Phi() || base_is_phi) && 1606 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) { 1607 return nullptr; // Neither memory or base are Phi 1608 } 1609 1610 if (mem->is_Phi()) { 1611 if (!stable_phi(mem->as_Phi(), phase)) { 1612 return nullptr; // Wait stable graph 1613 } 1614 uint cnt = mem->req(); 1615 // Check for loop invariant memory. 1616 if (cnt == 3) { 1617 for (uint i = 1; i < cnt; i++) { 1618 Node* in = mem->in(i); 1619 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1620 if (m == mem) { 1621 if (i == 1) { 1622 // if the first edge was a loop, check second edge too. 1623 // If both are replaceable - we are in an infinite loop 1624 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase); 1625 if (n == mem) { 1626 break; 1627 } 1628 } 1629 set_req(Memory, mem->in(cnt - i)); 1630 return this; // made change 1631 } 1632 } 1633 } 1634 } 1635 if (base_is_phi) { 1636 if (!stable_phi(base->as_Phi(), phase)) { 1637 return nullptr; // Wait stable graph 1638 } 1639 uint cnt = base->req(); 1640 // Check for loop invariant memory. 1641 if (cnt == 3) { 1642 for (uint i = 1; i < cnt; i++) { 1643 if (base->in(i) == base) { 1644 return nullptr; // Wait stable graph 1645 } 1646 } 1647 } 1648 } 1649 1650 // Split through Phi (see original code in loopopts.cpp). 1651 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type"); 1652 1653 // Do nothing here if Identity will find a value 1654 // (to avoid infinite chain of value phis generation). 1655 if (this != Identity(phase)) { 1656 return nullptr; 1657 } 1658 1659 // Select Region to split through. 1660 Node* region; 1661 if (!base_is_phi) { 1662 assert(mem->is_Phi(), "sanity"); 1663 region = mem->in(0); 1664 // Skip if the region dominates some control edge of the address. 1665 if (!MemNode::all_controls_dominate(address, region)) 1666 return nullptr; 1667 } else if (!mem->is_Phi()) { 1668 assert(base_is_phi, "sanity"); 1669 region = base->in(0); 1670 // Skip if the region dominates some control edge of the memory. 1671 if (!MemNode::all_controls_dominate(mem, region)) 1672 return nullptr; 1673 } else if (base->in(0) != mem->in(0)) { 1674 assert(base_is_phi && mem->is_Phi(), "sanity"); 1675 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1676 region = base->in(0); 1677 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1678 region = mem->in(0); 1679 } else { 1680 return nullptr; // complex graph 1681 } 1682 } else { 1683 assert(base->in(0) == mem->in(0), "sanity"); 1684 region = mem->in(0); 1685 } 1686 1687 Node* phi = nullptr; 1688 const Type* this_type = this->bottom_type(); 1689 PhaseIterGVN* igvn = phase->is_IterGVN(); 1690 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) { 1691 int this_index = C->get_alias_index(t_oop); 1692 int this_offset = t_oop->offset(); 1693 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx; 1694 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset); 1695 } else if (ignore_missing_instance_id) { 1696 phi = new PhiNode(region, this_type, nullptr, mem->_idx); 1697 } else { 1698 return nullptr; 1699 } 1700 1701 for (uint i = 1; i < region->req(); i++) { 1702 Node* x; 1703 Node* the_clone = nullptr; 1704 Node* in = region->in(i); 1705 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl && 1706 in != nullptr && in->is_OuterStripMinedLoop()) { 1707 // No node should go in the outer strip mined loop 1708 in = in->in(LoopNode::EntryControl); 1709 } 1710 if (in == nullptr || in == C->top()) { 1711 x = C->top(); // Dead path? Use a dead data op 1712 } else { 1713 x = this->clone(); // Else clone up the data op 1714 the_clone = x; // Remember for possible deletion. 1715 // Alter data node to use pre-phi inputs 1716 if (this->in(0) == region) { 1717 x->set_req(0, in); 1718 } else { 1719 x->set_req(0, nullptr); 1720 } 1721 if (mem->is_Phi() && (mem->in(0) == region)) { 1722 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1723 } 1724 if (address->is_Phi() && address->in(0) == region) { 1725 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1726 } 1727 if (base_is_phi && (base->in(0) == region)) { 1728 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1729 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1730 x->set_req(Address, adr_x); 1731 } 1732 } 1733 // Check for a 'win' on some paths 1734 const Type *t = x->Value(igvn); 1735 1736 bool singleton = t->singleton(); 1737 1738 // See comments in PhaseIdealLoop::split_thru_phi(). 1739 if (singleton && t == Type::TOP) { 1740 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1741 } 1742 1743 if (singleton) { 1744 x = igvn->makecon(t); 1745 } else { 1746 // We now call Identity to try to simplify the cloned node. 1747 // Note that some Identity methods call phase->type(this). 1748 // Make sure that the type array is big enough for 1749 // our new node, even though we may throw the node away. 1750 // (This tweaking with igvn only works because x is a new node.) 1751 igvn->set_type(x, t); 1752 // If x is a TypeNode, capture any more-precise type permanently into Node 1753 // otherwise it will be not updated during igvn->transform since 1754 // igvn->type(x) is set to x->Value() already. 1755 x->raise_bottom_type(t); 1756 Node* y = x->Identity(igvn); 1757 if (y != x) { 1758 x = y; 1759 } else { 1760 y = igvn->hash_find_insert(x); 1761 if (y) { 1762 x = y; 1763 } else { 1764 // Else x is a new node we are keeping 1765 // We do not need register_new_node_with_optimizer 1766 // because set_type has already been called. 1767 igvn->_worklist.push(x); 1768 } 1769 } 1770 } 1771 if (x != the_clone && the_clone != nullptr) { 1772 igvn->remove_dead_node(the_clone); 1773 } 1774 phi->set_req(i, x); 1775 } 1776 // Record Phi 1777 igvn->register_new_node_with_optimizer(phi); 1778 return phi; 1779 } 1780 1781 AllocateNode* LoadNode::is_new_object_mark_load() const { 1782 if (Opcode() == Op_LoadX) { 1783 Node* address = in(MemNode::Address); 1784 AllocateNode* alloc = AllocateNode::Ideal_allocation(address); 1785 Node* mem = in(MemNode::Memory); 1786 if (alloc != nullptr && mem->is_Proj() && 1787 mem->in(0) != nullptr && 1788 mem->in(0) == alloc->initialization() && 1789 alloc->initialization()->proj_out_or_null(0) != nullptr) { 1790 return alloc; 1791 } 1792 } 1793 return nullptr; 1794 } 1795 1796 1797 //------------------------------Ideal------------------------------------------ 1798 // If the load is from Field memory and the pointer is non-null, it might be possible to 1799 // zero out the control input. 1800 // If the offset is constant and the base is an object allocation, 1801 // try to hook me up to the exact initializing store. 1802 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1803 if (has_pinned_control_dependency()) { 1804 return nullptr; 1805 } 1806 Node* p = MemNode::Ideal_common(phase, can_reshape); 1807 if (p) return (p == NodeSentinel) ? nullptr : p; 1808 1809 Node* ctrl = in(MemNode::Control); 1810 Node* address = in(MemNode::Address); 1811 bool progress = false; 1812 1813 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) && 1814 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes()); 1815 1816 // Skip up past a SafePoint control. Cannot do this for Stores because 1817 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1818 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint && 1819 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw && 1820 !addr_mark && 1821 (depends_only_on_test() || has_unknown_control_dependency())) { 1822 ctrl = ctrl->in(0); 1823 set_req(MemNode::Control,ctrl); 1824 progress = true; 1825 } 1826 1827 intptr_t ignore = 0; 1828 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1829 if (base != nullptr 1830 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1831 // Check for useless control edge in some common special cases 1832 if (in(MemNode::Control) != nullptr 1833 && can_remove_control() 1834 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1835 && all_controls_dominate(base, phase->C->start())) { 1836 // A method-invariant, non-null address (constant or 'this' argument). 1837 set_req(MemNode::Control, nullptr); 1838 progress = true; 1839 } 1840 } 1841 1842 Node* mem = in(MemNode::Memory); 1843 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1844 1845 if (can_reshape && (addr_t != nullptr)) { 1846 // try to optimize our memory input 1847 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1848 if (opt_mem != mem) { 1849 set_req_X(MemNode::Memory, opt_mem, phase); 1850 if (phase->type( opt_mem ) == Type::TOP) return nullptr; 1851 return this; 1852 } 1853 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1854 if ((t_oop != nullptr) && 1855 (t_oop->is_known_instance_field() || 1856 t_oop->is_ptr_to_boxed_value())) { 1857 PhaseIterGVN *igvn = phase->is_IterGVN(); 1858 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true"); 1859 if (igvn->_worklist.member(opt_mem)) { 1860 // Delay this transformation until memory Phi is processed. 1861 igvn->_worklist.push(this); 1862 return nullptr; 1863 } 1864 // Split instance field load through Phi. 1865 Node* result = split_through_phi(phase); 1866 if (result != nullptr) return result; 1867 1868 if (t_oop->is_ptr_to_boxed_value()) { 1869 Node* result = eliminate_autobox(igvn); 1870 if (result != nullptr) return result; 1871 } 1872 } 1873 } 1874 1875 // Is there a dominating load that loads the same value? Leave 1876 // anything that is not a load of a field/array element (like 1877 // barriers etc.) alone 1878 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) { 1879 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 1880 Node *use = mem->fast_out(i); 1881 if (use != this && 1882 use->Opcode() == Opcode() && 1883 use->in(0) != nullptr && 1884 use->in(0) != in(0) && 1885 use->in(Address) == in(Address)) { 1886 Node* ctl = in(0); 1887 for (int i = 0; i < 10 && ctl != nullptr; i++) { 1888 ctl = IfNode::up_one_dom(ctl); 1889 if (ctl == use->in(0)) { 1890 set_req(0, use->in(0)); 1891 return this; 1892 } 1893 } 1894 } 1895 } 1896 } 1897 1898 // Check for prior store with a different base or offset; make Load 1899 // independent. Skip through any number of them. Bail out if the stores 1900 // are in an endless dead cycle and report no progress. This is a key 1901 // transform for Reflection. However, if after skipping through the Stores 1902 // we can't then fold up against a prior store do NOT do the transform as 1903 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1904 // array memory alive twice: once for the hoisted Load and again after the 1905 // bypassed Store. This situation only works if EVERYBODY who does 1906 // anti-dependence work knows how to bypass. I.e. we need all 1907 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1908 // the alias index stuff. So instead, peek through Stores and IFF we can 1909 // fold up, do so. 1910 Node* prev_mem = find_previous_store(phase); 1911 if (prev_mem != nullptr) { 1912 Node* value = can_see_arraycopy_value(prev_mem, phase); 1913 if (value != nullptr) { 1914 return value; 1915 } 1916 } 1917 // Steps (a), (b): Walk past independent stores to find an exact match. 1918 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) { 1919 // (c) See if we can fold up on the spot, but don't fold up here. 1920 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1921 // just return a prior value, which is done by Identity calls. 1922 if (can_see_stored_value(prev_mem, phase)) { 1923 // Make ready for step (d): 1924 set_req_X(MemNode::Memory, prev_mem, phase); 1925 return this; 1926 } 1927 } 1928 1929 return progress ? this : nullptr; 1930 } 1931 1932 // Helper to recognize certain Klass fields which are invariant across 1933 // some group of array types (e.g., int[] or all T[] where T < Object). 1934 const Type* 1935 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1936 ciKlass* klass) const { 1937 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1938 // The field is Klass::_modifier_flags. Return its (constant) value. 1939 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1940 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1941 return TypeInt::make(klass->modifier_flags()); 1942 } 1943 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1944 // The field is Klass::_access_flags. Return its (constant) value. 1945 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1946 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1947 return TypeInt::make(klass->access_flags()); 1948 } 1949 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1950 // The field is Klass::_layout_helper. Return its constant value if known. 1951 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1952 return TypeInt::make(klass->layout_helper()); 1953 } 1954 1955 // No match. 1956 return nullptr; 1957 } 1958 1959 //------------------------------Value----------------------------------------- 1960 const Type* LoadNode::Value(PhaseGVN* phase) const { 1961 // Either input is TOP ==> the result is TOP 1962 Node* mem = in(MemNode::Memory); 1963 const Type *t1 = phase->type(mem); 1964 if (t1 == Type::TOP) return Type::TOP; 1965 Node* adr = in(MemNode::Address); 1966 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1967 if (tp == nullptr || tp->empty()) return Type::TOP; 1968 int off = tp->offset(); 1969 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1970 Compile* C = phase->C; 1971 1972 // Try to guess loaded type from pointer type 1973 if (tp->isa_aryptr()) { 1974 const TypeAryPtr* ary = tp->is_aryptr(); 1975 const Type* t = ary->elem(); 1976 1977 // Determine whether the reference is beyond the header or not, by comparing 1978 // the offset against the offset of the start of the array's data. 1979 // Different array types begin at slightly different offsets (12 vs. 16). 1980 // We choose T_BYTE as an example base type that is least restrictive 1981 // as to alignment, which will therefore produce the smallest 1982 // possible base offset. 1983 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1984 const bool off_beyond_header = (off >= min_base_off); 1985 1986 // Try to constant-fold a stable array element. 1987 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) { 1988 // Make sure the reference is not into the header and the offset is constant 1989 ciObject* aobj = ary->const_oop(); 1990 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1991 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0); 1992 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, 1993 stable_dimension, 1994 memory_type(), is_unsigned()); 1995 if (con_type != nullptr) { 1996 return con_type; 1997 } 1998 } 1999 } 2000 2001 // Don't do this for integer types. There is only potential profit if 2002 // the element type t is lower than _type; that is, for int types, if _type is 2003 // more restrictive than t. This only happens here if one is short and the other 2004 // char (both 16 bits), and in those cases we've made an intentional decision 2005 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 2006 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 2007 // 2008 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 2009 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 2010 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 2011 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 2012 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 2013 // In fact, that could have been the original type of p1, and p1 could have 2014 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 2015 // expression (LShiftL quux 3) independently optimized to the constant 8. 2016 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr) 2017 && (_type->isa_vect() == nullptr) 2018 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 2019 // t might actually be lower than _type, if _type is a unique 2020 // concrete subclass of abstract class t. 2021 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header? 2022 const Type* jt = t->join_speculative(_type); 2023 // In any case, do not allow the join, per se, to empty out the type. 2024 if (jt->empty() && !t->empty()) { 2025 // This can happen if a interface-typed array narrows to a class type. 2026 jt = _type; 2027 } 2028 #ifdef ASSERT 2029 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 2030 // The pointers in the autobox arrays are always non-null 2031 Node* base = adr->in(AddPNode::Base); 2032 if ((base != nullptr) && base->is_DecodeN()) { 2033 // Get LoadN node which loads IntegerCache.cache field 2034 base = base->in(1); 2035 } 2036 if ((base != nullptr) && base->is_Con()) { 2037 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 2038 if ((base_type != nullptr) && base_type->is_autobox_cache()) { 2039 // It could be narrow oop 2040 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 2041 } 2042 } 2043 } 2044 #endif 2045 return jt; 2046 } 2047 } 2048 } else if (tp->base() == Type::InstPtr) { 2049 assert( off != Type::OffsetBot || 2050 // arrays can be cast to Objects 2051 !tp->isa_instptr() || 2052 tp->is_instptr()->instance_klass()->is_java_lang_Object() || 2053 // unsafe field access may not have a constant offset 2054 C->has_unsafe_access(), 2055 "Field accesses must be precise" ); 2056 // For oop loads, we expect the _type to be precise. 2057 2058 // Optimize loads from constant fields. 2059 const TypeInstPtr* tinst = tp->is_instptr(); 2060 ciObject* const_oop = tinst->const_oop(); 2061 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) { 2062 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type()); 2063 if (con_type != nullptr) { 2064 return con_type; 2065 } 2066 } 2067 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) { 2068 assert(off != Type::OffsetBot || 2069 !tp->isa_instklassptr() || 2070 // arrays can be cast to Objects 2071 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() || 2072 // also allow array-loading from the primary supertype 2073 // array during subtype checks 2074 Opcode() == Op_LoadKlass, 2075 "Field accesses must be precise"); 2076 // For klass/static loads, we expect the _type to be precise 2077 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) { 2078 /* With mirrors being an indirect in the Klass* 2079 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset)) 2080 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass). 2081 * 2082 * So check the type and klass of the node before the LoadP. 2083 */ 2084 Node* adr2 = adr->in(MemNode::Address); 2085 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2086 if (tkls != nullptr && !StressReflectiveCode) { 2087 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 2088 ciKlass* klass = tkls->exact_klass(); 2089 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 2090 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 2091 return TypeInstPtr::make(klass->java_mirror()); 2092 } 2093 } 2094 } 2095 2096 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2097 if (tkls != nullptr) { 2098 if (tkls->is_loaded() && tkls->klass_is_exact()) { 2099 ciKlass* klass = tkls->exact_klass(); 2100 // We are loading a field from a Klass metaobject whose identity 2101 // is known at compile time (the type is "exact" or "precise"). 2102 // Check for fields we know are maintained as constants by the VM. 2103 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 2104 // The field is Klass::_super_check_offset. Return its (constant) value. 2105 // (Folds up type checking code.) 2106 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 2107 return TypeInt::make(klass->super_check_offset()); 2108 } 2109 // Compute index into primary_supers array 2110 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 2111 // Check for overflowing; use unsigned compare to handle the negative case. 2112 if( depth < ciKlass::primary_super_limit() ) { 2113 // The field is an element of Klass::_primary_supers. Return its (constant) value. 2114 // (Folds up type checking code.) 2115 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 2116 ciKlass *ss = klass->super_of_depth(depth); 2117 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR; 2118 } 2119 const Type* aift = load_array_final_field(tkls, klass); 2120 if (aift != nullptr) return aift; 2121 } 2122 2123 // We can still check if we are loading from the primary_supers array at a 2124 // shallow enough depth. Even though the klass is not exact, entries less 2125 // than or equal to its super depth are correct. 2126 if (tkls->is_loaded()) { 2127 ciKlass* klass = nullptr; 2128 if (tkls->isa_instklassptr()) { 2129 klass = tkls->is_instklassptr()->instance_klass(); 2130 } else { 2131 int dims; 2132 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims); 2133 if (inner->isa_instklassptr()) { 2134 klass = inner->is_instklassptr()->instance_klass(); 2135 klass = ciObjArrayKlass::make(klass, dims); 2136 } 2137 } 2138 if (klass != nullptr) { 2139 // Compute index into primary_supers array 2140 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 2141 // Check for overflowing; use unsigned compare to handle the negative case. 2142 if (depth < ciKlass::primary_super_limit() && 2143 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case 2144 // The field is an element of Klass::_primary_supers. Return its (constant) value. 2145 // (Folds up type checking code.) 2146 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 2147 ciKlass *ss = klass->super_of_depth(depth); 2148 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR; 2149 } 2150 } 2151 } 2152 2153 // If the type is enough to determine that the thing is not an array, 2154 // we can give the layout_helper a positive interval type. 2155 // This will help short-circuit some reflective code. 2156 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) && 2157 tkls->isa_instklassptr() && // not directly typed as an array 2158 !tkls->is_instklassptr()->instance_klass()->is_java_lang_Object() // not the supertype of all T[] and specifically not Serializable & Cloneable 2159 ) { 2160 // Note: When interfaces are reliable, we can narrow the interface 2161 // test to (klass != Serializable && klass != Cloneable). 2162 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 2163 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 2164 // The key property of this type is that it folds up tests 2165 // for array-ness, since it proves that the layout_helper is positive. 2166 // Thus, a generic value like the basic object layout helper works fine. 2167 return TypeInt::make(min_size, max_jint, Type::WidenMin); 2168 } 2169 } 2170 2171 // If we are loading from a freshly-allocated object, produce a zero, 2172 // if the load is provably beyond the header of the object. 2173 // (Also allow a variable load from a fresh array to produce zero.) 2174 const TypeOopPtr *tinst = tp->isa_oopptr(); 2175 bool is_instance = (tinst != nullptr) && tinst->is_known_instance_field(); 2176 bool is_boxed_value = (tinst != nullptr) && tinst->is_ptr_to_boxed_value(); 2177 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 2178 Node* value = can_see_stored_value(mem,phase); 2179 if (value != nullptr && value->is_Con()) { 2180 assert(value->bottom_type()->higher_equal(_type),"sanity"); 2181 return value->bottom_type(); 2182 } 2183 } 2184 2185 bool is_vect = (_type->isa_vect() != nullptr); 2186 if (is_instance && !is_vect) { 2187 // If we have an instance type and our memory input is the 2188 // programs's initial memory state, there is no matching store, 2189 // so just return a zero of the appropriate type - 2190 // except if it is vectorized - then we have no zero constant. 2191 Node *mem = in(MemNode::Memory); 2192 if (mem->is_Parm() && mem->in(0)->is_Start()) { 2193 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 2194 return Type::get_zero_type(_type->basic_type()); 2195 } 2196 } 2197 2198 Node* alloc = is_new_object_mark_load(); 2199 if (alloc != nullptr) { 2200 return TypeX::make(markWord::prototype().value()); 2201 } 2202 2203 return _type; 2204 } 2205 2206 //------------------------------match_edge------------------------------------- 2207 // Do we Match on this edge index or not? Match only the address. 2208 uint LoadNode::match_edge(uint idx) const { 2209 return idx == MemNode::Address; 2210 } 2211 2212 //--------------------------LoadBNode::Ideal-------------------------------------- 2213 // 2214 // If the previous store is to the same address as this load, 2215 // and the value stored was larger than a byte, replace this load 2216 // with the value stored truncated to a byte. If no truncation is 2217 // needed, the replacement is done in LoadNode::Identity(). 2218 // 2219 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2220 Node* mem = in(MemNode::Memory); 2221 Node* value = can_see_stored_value(mem,phase); 2222 if (value != nullptr) { 2223 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false); 2224 if (narrow != value) { 2225 return narrow; 2226 } 2227 } 2228 // Identity call will handle the case where truncation is not needed. 2229 return LoadNode::Ideal(phase, can_reshape); 2230 } 2231 2232 const Type* LoadBNode::Value(PhaseGVN* phase) const { 2233 Node* mem = in(MemNode::Memory); 2234 Node* value = can_see_stored_value(mem,phase); 2235 if (value != nullptr && value->is_Con() && 2236 !value->bottom_type()->higher_equal(_type)) { 2237 // If the input to the store does not fit with the load's result type, 2238 // it must be truncated. We can't delay until Ideal call since 2239 // a singleton Value is needed for split_thru_phi optimization. 2240 int con = value->get_int(); 2241 return TypeInt::make((con << 24) >> 24); 2242 } 2243 return LoadNode::Value(phase); 2244 } 2245 2246 //--------------------------LoadUBNode::Ideal------------------------------------- 2247 // 2248 // If the previous store is to the same address as this load, 2249 // and the value stored was larger than a byte, replace this load 2250 // with the value stored truncated to a byte. If no truncation is 2251 // needed, the replacement is done in LoadNode::Identity(). 2252 // 2253 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2254 Node* mem = in(MemNode::Memory); 2255 Node* value = can_see_stored_value(mem, phase); 2256 if (value != nullptr) { 2257 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false); 2258 if (narrow != value) { 2259 return narrow; 2260 } 2261 } 2262 // Identity call will handle the case where truncation is not needed. 2263 return LoadNode::Ideal(phase, can_reshape); 2264 } 2265 2266 const Type* LoadUBNode::Value(PhaseGVN* phase) const { 2267 Node* mem = in(MemNode::Memory); 2268 Node* value = can_see_stored_value(mem,phase); 2269 if (value != nullptr && value->is_Con() && 2270 !value->bottom_type()->higher_equal(_type)) { 2271 // If the input to the store does not fit with the load's result type, 2272 // it must be truncated. We can't delay until Ideal call since 2273 // a singleton Value is needed for split_thru_phi optimization. 2274 int con = value->get_int(); 2275 return TypeInt::make(con & 0xFF); 2276 } 2277 return LoadNode::Value(phase); 2278 } 2279 2280 //--------------------------LoadUSNode::Ideal------------------------------------- 2281 // 2282 // If the previous store is to the same address as this load, 2283 // and the value stored was larger than a char, replace this load 2284 // with the value stored truncated to a char. If no truncation is 2285 // needed, the replacement is done in LoadNode::Identity(). 2286 // 2287 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2288 Node* mem = in(MemNode::Memory); 2289 Node* value = can_see_stored_value(mem,phase); 2290 if (value != nullptr) { 2291 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false); 2292 if (narrow != value) { 2293 return narrow; 2294 } 2295 } 2296 // Identity call will handle the case where truncation is not needed. 2297 return LoadNode::Ideal(phase, can_reshape); 2298 } 2299 2300 const Type* LoadUSNode::Value(PhaseGVN* phase) const { 2301 Node* mem = in(MemNode::Memory); 2302 Node* value = can_see_stored_value(mem,phase); 2303 if (value != nullptr && value->is_Con() && 2304 !value->bottom_type()->higher_equal(_type)) { 2305 // If the input to the store does not fit with the load's result type, 2306 // it must be truncated. We can't delay until Ideal call since 2307 // a singleton Value is needed for split_thru_phi optimization. 2308 int con = value->get_int(); 2309 return TypeInt::make(con & 0xFFFF); 2310 } 2311 return LoadNode::Value(phase); 2312 } 2313 2314 //--------------------------LoadSNode::Ideal-------------------------------------- 2315 // 2316 // If the previous store is to the same address as this load, 2317 // and the value stored was larger than a short, replace this load 2318 // with the value stored truncated to a short. If no truncation is 2319 // needed, the replacement is done in LoadNode::Identity(). 2320 // 2321 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) { 2322 Node* mem = in(MemNode::Memory); 2323 Node* value = can_see_stored_value(mem,phase); 2324 if (value != nullptr) { 2325 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false); 2326 if (narrow != value) { 2327 return narrow; 2328 } 2329 } 2330 // Identity call will handle the case where truncation is not needed. 2331 return LoadNode::Ideal(phase, can_reshape); 2332 } 2333 2334 const Type* LoadSNode::Value(PhaseGVN* phase) const { 2335 Node* mem = in(MemNode::Memory); 2336 Node* value = can_see_stored_value(mem,phase); 2337 if (value != nullptr && value->is_Con() && 2338 !value->bottom_type()->higher_equal(_type)) { 2339 // If the input to the store does not fit with the load's result type, 2340 // it must be truncated. We can't delay until Ideal call since 2341 // a singleton Value is needed for split_thru_phi optimization. 2342 int con = value->get_int(); 2343 return TypeInt::make((con << 16) >> 16); 2344 } 2345 return LoadNode::Value(phase); 2346 } 2347 2348 //============================================================================= 2349 //----------------------------LoadKlassNode::make------------------------------ 2350 // Polymorphic factory method: 2351 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) { 2352 // sanity check the alias category against the created node type 2353 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2354 assert(adr_type != nullptr, "expecting TypeKlassPtr"); 2355 #ifdef _LP64 2356 if (adr_type->is_ptr_to_narrowklass()) { 2357 assert(UseCompressedClassPointers, "no compressed klasses"); 2358 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2359 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2360 } 2361 #endif 2362 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2363 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 2364 } 2365 2366 //------------------------------Value------------------------------------------ 2367 const Type* LoadKlassNode::Value(PhaseGVN* phase) const { 2368 return klass_value_common(phase); 2369 } 2370 2371 // In most cases, LoadKlassNode does not have the control input set. If the control 2372 // input is set, it must not be removed (by LoadNode::Ideal()). 2373 bool LoadKlassNode::can_remove_control() const { 2374 return false; 2375 } 2376 2377 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const { 2378 // Either input is TOP ==> the result is TOP 2379 const Type *t1 = phase->type( in(MemNode::Memory) ); 2380 if (t1 == Type::TOP) return Type::TOP; 2381 Node *adr = in(MemNode::Address); 2382 const Type *t2 = phase->type( adr ); 2383 if (t2 == Type::TOP) return Type::TOP; 2384 const TypePtr *tp = t2->is_ptr(); 2385 if (TypePtr::above_centerline(tp->ptr()) || 2386 tp->ptr() == TypePtr::Null) return Type::TOP; 2387 2388 // Return a more precise klass, if possible 2389 const TypeInstPtr *tinst = tp->isa_instptr(); 2390 if (tinst != nullptr) { 2391 ciInstanceKlass* ik = tinst->instance_klass(); 2392 int offset = tinst->offset(); 2393 if (ik == phase->C->env()->Class_klass() 2394 && (offset == java_lang_Class::klass_offset() || 2395 offset == java_lang_Class::array_klass_offset())) { 2396 // We are loading a special hidden field from a Class mirror object, 2397 // the field which points to the VM's Klass metaobject. 2398 ciType* t = tinst->java_mirror_type(); 2399 // java_mirror_type returns non-null for compile-time Class constants. 2400 if (t != nullptr) { 2401 // constant oop => constant klass 2402 if (offset == java_lang_Class::array_klass_offset()) { 2403 if (t->is_void()) { 2404 // We cannot create a void array. Since void is a primitive type return null 2405 // klass. Users of this result need to do a null check on the returned klass. 2406 return TypePtr::NULL_PTR; 2407 } 2408 return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces); 2409 } 2410 if (!t->is_klass()) { 2411 // a primitive Class (e.g., int.class) has null for a klass field 2412 return TypePtr::NULL_PTR; 2413 } 2414 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2415 return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces); 2416 } 2417 // non-constant mirror, so we can't tell what's going on 2418 } 2419 if (!tinst->is_loaded()) 2420 return _type; // Bail out if not loaded 2421 if (offset == oopDesc::klass_offset_in_bytes()) { 2422 return tinst->as_klass_type(true); 2423 } 2424 } 2425 2426 // Check for loading klass from an array 2427 const TypeAryPtr *tary = tp->isa_aryptr(); 2428 if (tary != nullptr && 2429 tary->offset() == oopDesc::klass_offset_in_bytes()) { 2430 return tary->as_klass_type(true); 2431 } 2432 2433 // Check for loading klass from an array klass 2434 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2435 if (tkls != nullptr && !StressReflectiveCode) { 2436 if (!tkls->is_loaded()) 2437 return _type; // Bail out if not loaded 2438 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() && 2439 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2440 // // Always returning precise element type is incorrect, 2441 // // e.g., element type could be object and array may contain strings 2442 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2443 2444 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2445 // according to the element type's subclassing. 2446 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact()); 2447 } 2448 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() && 2449 tkls->offset() == in_bytes(Klass::super_offset())) { 2450 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super(); 2451 // The field is Klass::_super. Return its (constant) value. 2452 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2453 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR; 2454 } 2455 } 2456 2457 if (tkls != nullptr && !UseSecondarySupersCache 2458 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) { 2459 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled. 2460 return TypePtr::NULL_PTR; 2461 } 2462 2463 // Bailout case 2464 return LoadNode::Value(phase); 2465 } 2466 2467 //------------------------------Identity--------------------------------------- 2468 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2469 // Also feed through the klass in Allocate(...klass...)._klass. 2470 Node* LoadKlassNode::Identity(PhaseGVN* phase) { 2471 return klass_identity_common(phase); 2472 } 2473 2474 Node* LoadNode::klass_identity_common(PhaseGVN* phase) { 2475 Node* x = LoadNode::Identity(phase); 2476 if (x != this) return x; 2477 2478 // Take apart the address into an oop and offset. 2479 // Return 'this' if we cannot. 2480 Node* adr = in(MemNode::Address); 2481 intptr_t offset = 0; 2482 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2483 if (base == nullptr) return this; 2484 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2485 if (toop == nullptr) return this; 2486 2487 // Step over potential GC barrier for OopHandle resolve 2488 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2489 if (bs->is_gc_barrier_node(base)) { 2490 base = bs->step_over_gc_barrier(base); 2491 } 2492 2493 // We can fetch the klass directly through an AllocateNode. 2494 // This works even if the klass is not constant (clone or newArray). 2495 if (offset == oopDesc::klass_offset_in_bytes()) { 2496 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2497 if (allocated_klass != nullptr) { 2498 return allocated_klass; 2499 } 2500 } 2501 2502 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2503 // See inline_native_Class_query for occurrences of these patterns. 2504 // Java Example: x.getClass().isAssignableFrom(y) 2505 // 2506 // This improves reflective code, often making the Class 2507 // mirror go completely dead. (Current exception: Class 2508 // mirrors may appear in debug info, but we could clean them out by 2509 // introducing a new debug info operator for Klass.java_mirror). 2510 2511 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass() 2512 && offset == java_lang_Class::klass_offset()) { 2513 if (base->is_Load()) { 2514 Node* base2 = base->in(MemNode::Address); 2515 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */ 2516 Node* adr2 = base2->in(MemNode::Address); 2517 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2518 if (tkls != nullptr && !tkls->empty() 2519 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr()) 2520 && adr2->is_AddP() 2521 ) { 2522 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2523 if (tkls->offset() == mirror_field) { 2524 return adr2->in(AddPNode::Base); 2525 } 2526 } 2527 } 2528 } 2529 } 2530 2531 return this; 2532 } 2533 2534 LoadNode* LoadNode::clone_pinned() const { 2535 LoadNode* ld = clone()->as_Load(); 2536 ld->_control_dependency = UnknownControl; 2537 return ld; 2538 } 2539 2540 2541 //------------------------------Value------------------------------------------ 2542 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const { 2543 const Type *t = klass_value_common(phase); 2544 if (t == Type::TOP) 2545 return t; 2546 2547 return t->make_narrowklass(); 2548 } 2549 2550 //------------------------------Identity--------------------------------------- 2551 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2552 // Also feed through the klass in Allocate(...klass...)._klass. 2553 Node* LoadNKlassNode::Identity(PhaseGVN* phase) { 2554 Node *x = klass_identity_common(phase); 2555 2556 const Type *t = phase->type( x ); 2557 if( t == Type::TOP ) return x; 2558 if( t->isa_narrowklass()) return x; 2559 assert (!t->isa_narrowoop(), "no narrow oop here"); 2560 2561 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2562 } 2563 2564 //------------------------------Value----------------------------------------- 2565 const Type* LoadRangeNode::Value(PhaseGVN* phase) const { 2566 // Either input is TOP ==> the result is TOP 2567 const Type *t1 = phase->type( in(MemNode::Memory) ); 2568 if( t1 == Type::TOP ) return Type::TOP; 2569 Node *adr = in(MemNode::Address); 2570 const Type *t2 = phase->type( adr ); 2571 if( t2 == Type::TOP ) return Type::TOP; 2572 const TypePtr *tp = t2->is_ptr(); 2573 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2574 const TypeAryPtr *tap = tp->isa_aryptr(); 2575 if( !tap ) return _type; 2576 return tap->size(); 2577 } 2578 2579 //-------------------------------Ideal--------------------------------------- 2580 // Feed through the length in AllocateArray(...length...)._length. 2581 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2582 Node* p = MemNode::Ideal_common(phase, can_reshape); 2583 if (p) return (p == NodeSentinel) ? nullptr : p; 2584 2585 // Take apart the address into an oop and offset. 2586 // Return 'this' if we cannot. 2587 Node* adr = in(MemNode::Address); 2588 intptr_t offset = 0; 2589 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2590 if (base == nullptr) return nullptr; 2591 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2592 if (tary == nullptr) return nullptr; 2593 2594 // We can fetch the length directly through an AllocateArrayNode. 2595 // This works even if the length is not constant (clone or newArray). 2596 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2597 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base); 2598 if (alloc != nullptr) { 2599 Node* allocated_length = alloc->Ideal_length(); 2600 Node* len = alloc->make_ideal_length(tary, phase); 2601 if (allocated_length != len) { 2602 // New CastII improves on this. 2603 return len; 2604 } 2605 } 2606 } 2607 2608 return nullptr; 2609 } 2610 2611 //------------------------------Identity--------------------------------------- 2612 // Feed through the length in AllocateArray(...length...)._length. 2613 Node* LoadRangeNode::Identity(PhaseGVN* phase) { 2614 Node* x = LoadINode::Identity(phase); 2615 if (x != this) return x; 2616 2617 // Take apart the address into an oop and offset. 2618 // Return 'this' if we cannot. 2619 Node* adr = in(MemNode::Address); 2620 intptr_t offset = 0; 2621 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2622 if (base == nullptr) return this; 2623 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2624 if (tary == nullptr) return this; 2625 2626 // We can fetch the length directly through an AllocateArrayNode. 2627 // This works even if the length is not constant (clone or newArray). 2628 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2629 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base); 2630 if (alloc != nullptr) { 2631 Node* allocated_length = alloc->Ideal_length(); 2632 // Do not allow make_ideal_length to allocate a CastII node. 2633 Node* len = alloc->make_ideal_length(tary, phase, false); 2634 if (allocated_length == len) { 2635 // Return allocated_length only if it would not be improved by a CastII. 2636 return allocated_length; 2637 } 2638 } 2639 } 2640 2641 return this; 2642 2643 } 2644 2645 //============================================================================= 2646 //---------------------------StoreNode::make----------------------------------- 2647 // Polymorphic factory method: 2648 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) { 2649 assert((mo == unordered || mo == release), "unexpected"); 2650 Compile* C = gvn.C; 2651 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2652 ctl != nullptr, "raw memory operations should have control edge"); 2653 2654 switch (bt) { 2655 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case 2656 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2657 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2658 case T_CHAR: 2659 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2660 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access); 2661 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2662 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access); 2663 case T_METADATA: 2664 case T_ADDRESS: 2665 case T_OBJECT: 2666 #ifdef _LP64 2667 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2668 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2669 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2670 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2671 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2672 adr->bottom_type()->isa_rawptr())) { 2673 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2674 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2675 } 2676 #endif 2677 { 2678 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2679 } 2680 default: 2681 ShouldNotReachHere(); 2682 return (StoreNode*)nullptr; 2683 } 2684 } 2685 2686 //--------------------------bottom_type---------------------------------------- 2687 const Type *StoreNode::bottom_type() const { 2688 return Type::MEMORY; 2689 } 2690 2691 //------------------------------hash------------------------------------------- 2692 uint StoreNode::hash() const { 2693 // unroll addition of interesting fields 2694 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2695 2696 // Since they are not commoned, do not hash them: 2697 return NO_HASH; 2698 } 2699 2700 //------------------------------Ideal------------------------------------------ 2701 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2702 // When a store immediately follows a relevant allocation/initialization, 2703 // try to capture it into the initialization, or hoist it above. 2704 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2705 Node* p = MemNode::Ideal_common(phase, can_reshape); 2706 if (p) return (p == NodeSentinel) ? nullptr : p; 2707 2708 Node* mem = in(MemNode::Memory); 2709 Node* address = in(MemNode::Address); 2710 Node* value = in(MemNode::ValueIn); 2711 // Back-to-back stores to same address? Fold em up. Generally 2712 // unsafe if I have intervening uses... Also disallowed for StoreCM 2713 // since they must follow each StoreP operation. Redundant StoreCMs 2714 // are eliminated just before matching in final_graph_reshape. 2715 { 2716 Node* st = mem; 2717 // If Store 'st' has more than one use, we cannot fold 'st' away. 2718 // For example, 'st' might be the final state at a conditional 2719 // return. Or, 'st' might be used by some node which is live at 2720 // the same time 'st' is live, which might be unschedulable. So, 2721 // require exactly ONE user until such time as we clone 'mem' for 2722 // each of 'mem's uses (thus making the exactly-1-user-rule hold 2723 // true). 2724 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) { 2725 // Looking at a dead closed cycle of memory? 2726 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2727 assert(Opcode() == st->Opcode() || 2728 st->Opcode() == Op_StoreVector || 2729 Opcode() == Op_StoreVector || 2730 st->Opcode() == Op_StoreVectorScatter || 2731 Opcode() == Op_StoreVectorScatter || 2732 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw || 2733 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode 2734 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy 2735 (is_mismatched_access() || st->as_Store()->is_mismatched_access()), 2736 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]); 2737 2738 if (st->in(MemNode::Address)->eqv_uncast(address) && 2739 st->as_Store()->memory_size() <= this->memory_size()) { 2740 Node* use = st->raw_out(0); 2741 if (phase->is_IterGVN()) { 2742 phase->is_IterGVN()->rehash_node_delayed(use); 2743 } 2744 // It's OK to do this in the parser, since DU info is always accurate, 2745 // and the parser always refers to nodes via SafePointNode maps. 2746 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase); 2747 return this; 2748 } 2749 st = st->in(MemNode::Memory); 2750 } 2751 } 2752 2753 2754 // Capture an unaliased, unconditional, simple store into an initializer. 2755 // Or, if it is independent of the allocation, hoist it above the allocation. 2756 if (ReduceFieldZeroing && /*can_reshape &&*/ 2757 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2758 InitializeNode* init = mem->in(0)->as_Initialize(); 2759 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2760 if (offset > 0) { 2761 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2762 // If the InitializeNode captured me, it made a raw copy of me, 2763 // and I need to disappear. 2764 if (moved != nullptr) { 2765 // %%% hack to ensure that Ideal returns a new node: 2766 mem = MergeMemNode::make(mem); 2767 return mem; // fold me away 2768 } 2769 } 2770 } 2771 2772 // Fold reinterpret cast into memory operation: 2773 // StoreX mem (MoveY2X v) => StoreY mem v 2774 if (value->is_Move()) { 2775 const Type* vt = value->in(1)->bottom_type(); 2776 if (has_reinterpret_variant(vt)) { 2777 if (phase->C->post_loop_opts_phase()) { 2778 return convert_to_reinterpret_store(*phase, value->in(1), vt); 2779 } else { 2780 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over 2781 } 2782 } 2783 } 2784 2785 return nullptr; // No further progress 2786 } 2787 2788 //------------------------------Value----------------------------------------- 2789 const Type* StoreNode::Value(PhaseGVN* phase) const { 2790 // Either input is TOP ==> the result is TOP 2791 const Type *t1 = phase->type( in(MemNode::Memory) ); 2792 if( t1 == Type::TOP ) return Type::TOP; 2793 const Type *t2 = phase->type( in(MemNode::Address) ); 2794 if( t2 == Type::TOP ) return Type::TOP; 2795 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2796 if( t3 == Type::TOP ) return Type::TOP; 2797 return Type::MEMORY; 2798 } 2799 2800 //------------------------------Identity--------------------------------------- 2801 // Remove redundant stores: 2802 // Store(m, p, Load(m, p)) changes to m. 2803 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2804 Node* StoreNode::Identity(PhaseGVN* phase) { 2805 Node* mem = in(MemNode::Memory); 2806 Node* adr = in(MemNode::Address); 2807 Node* val = in(MemNode::ValueIn); 2808 2809 Node* result = this; 2810 2811 // Load then Store? Then the Store is useless 2812 if (val->is_Load() && 2813 val->in(MemNode::Address)->eqv_uncast(adr) && 2814 val->in(MemNode::Memory )->eqv_uncast(mem) && 2815 val->as_Load()->store_Opcode() == Opcode()) { 2816 result = mem; 2817 } 2818 2819 // Two stores in a row of the same value? 2820 if (result == this && 2821 mem->is_Store() && 2822 mem->in(MemNode::Address)->eqv_uncast(adr) && 2823 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2824 mem->Opcode() == Opcode()) { 2825 result = mem; 2826 } 2827 2828 // Store of zero anywhere into a freshly-allocated object? 2829 // Then the store is useless. 2830 // (It must already have been captured by the InitializeNode.) 2831 if (result == this && 2832 ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2833 // a newly allocated object is already all-zeroes everywhere 2834 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2835 result = mem; 2836 } 2837 2838 if (result == this) { 2839 // the store may also apply to zero-bits in an earlier object 2840 Node* prev_mem = find_previous_store(phase); 2841 // Steps (a), (b): Walk past independent stores to find an exact match. 2842 if (prev_mem != nullptr) { 2843 Node* prev_val = can_see_stored_value(prev_mem, phase); 2844 if (prev_val != nullptr && prev_val == val) { 2845 // prev_val and val might differ by a cast; it would be good 2846 // to keep the more informative of the two. 2847 result = mem; 2848 } 2849 } 2850 } 2851 } 2852 2853 PhaseIterGVN* igvn = phase->is_IterGVN(); 2854 if (result != this && igvn != nullptr) { 2855 MemBarNode* trailing = trailing_membar(); 2856 if (trailing != nullptr) { 2857 #ifdef ASSERT 2858 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr(); 2859 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects"); 2860 #endif 2861 trailing->remove(igvn); 2862 } 2863 } 2864 2865 return result; 2866 } 2867 2868 //------------------------------match_edge------------------------------------- 2869 // Do we Match on this edge index or not? Match only memory & value 2870 uint StoreNode::match_edge(uint idx) const { 2871 return idx == MemNode::Address || idx == MemNode::ValueIn; 2872 } 2873 2874 //------------------------------cmp-------------------------------------------- 2875 // Do not common stores up together. They generally have to be split 2876 // back up anyways, so do not bother. 2877 bool StoreNode::cmp( const Node &n ) const { 2878 return (&n == this); // Always fail except on self 2879 } 2880 2881 //------------------------------Ideal_masked_input----------------------------- 2882 // Check for a useless mask before a partial-word store 2883 // (StoreB ... (AndI valIn conIa) ) 2884 // If (conIa & mask == mask) this simplifies to 2885 // (StoreB ... (valIn) ) 2886 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2887 Node *val = in(MemNode::ValueIn); 2888 if( val->Opcode() == Op_AndI ) { 2889 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2890 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2891 set_req_X(MemNode::ValueIn, val->in(1), phase); 2892 return this; 2893 } 2894 } 2895 return nullptr; 2896 } 2897 2898 2899 //------------------------------Ideal_sign_extended_input---------------------- 2900 // Check for useless sign-extension before a partial-word store 2901 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2902 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2903 // (StoreB ... (valIn) ) 2904 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2905 Node *val = in(MemNode::ValueIn); 2906 if( val->Opcode() == Op_RShiftI ) { 2907 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2908 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2909 Node *shl = val->in(1); 2910 if( shl->Opcode() == Op_LShiftI ) { 2911 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2912 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2913 set_req_X(MemNode::ValueIn, shl->in(1), phase); 2914 return this; 2915 } 2916 } 2917 } 2918 } 2919 return nullptr; 2920 } 2921 2922 //------------------------------value_never_loaded----------------------------------- 2923 // Determine whether there are any possible loads of the value stored. 2924 // For simplicity, we actually check if there are any loads from the 2925 // address stored to, not just for loads of the value stored by this node. 2926 // 2927 bool StoreNode::value_never_loaded(PhaseValues* phase) const { 2928 Node *adr = in(Address); 2929 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2930 if (adr_oop == nullptr) 2931 return false; 2932 if (!adr_oop->is_known_instance_field()) 2933 return false; // if not a distinct instance, there may be aliases of the address 2934 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2935 Node *use = adr->fast_out(i); 2936 if (use->is_Load() || use->is_LoadStore()) { 2937 return false; 2938 } 2939 } 2940 return true; 2941 } 2942 2943 MemBarNode* StoreNode::trailing_membar() const { 2944 if (is_release()) { 2945 MemBarNode* trailing_mb = nullptr; 2946 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 2947 Node* u = fast_out(i); 2948 if (u->is_MemBar()) { 2949 if (u->as_MemBar()->trailing_store()) { 2950 assert(u->Opcode() == Op_MemBarVolatile, ""); 2951 assert(trailing_mb == nullptr, "only one"); 2952 trailing_mb = u->as_MemBar(); 2953 #ifdef ASSERT 2954 Node* leading = u->as_MemBar()->leading_membar(); 2955 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 2956 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair"); 2957 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair"); 2958 #endif 2959 } else { 2960 assert(u->as_MemBar()->standalone(), ""); 2961 } 2962 } 2963 } 2964 return trailing_mb; 2965 } 2966 return nullptr; 2967 } 2968 2969 2970 //============================================================================= 2971 //------------------------------Ideal------------------------------------------ 2972 // If the store is from an AND mask that leaves the low bits untouched, then 2973 // we can skip the AND operation. If the store is from a sign-extension 2974 // (a left shift, then right shift) we can skip both. 2975 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2976 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2977 if( progress != nullptr ) return progress; 2978 2979 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2980 if( progress != nullptr ) return progress; 2981 2982 // Finally check the default case 2983 return StoreNode::Ideal(phase, can_reshape); 2984 } 2985 2986 //============================================================================= 2987 //------------------------------Ideal------------------------------------------ 2988 // If the store is from an AND mask that leaves the low bits untouched, then 2989 // we can skip the AND operation 2990 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2991 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2992 if( progress != nullptr ) return progress; 2993 2994 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2995 if( progress != nullptr ) return progress; 2996 2997 // Finally check the default case 2998 return StoreNode::Ideal(phase, can_reshape); 2999 } 3000 3001 //============================================================================= 3002 //------------------------------Identity--------------------------------------- 3003 Node* StoreCMNode::Identity(PhaseGVN* phase) { 3004 // No need to card mark when storing a null ptr 3005 Node* my_store = in(MemNode::OopStore); 3006 if (my_store->is_Store()) { 3007 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 3008 if( t1 == TypePtr::NULL_PTR ) { 3009 return in(MemNode::Memory); 3010 } 3011 } 3012 return this; 3013 } 3014 3015 //============================================================================= 3016 //------------------------------Ideal--------------------------------------- 3017 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 3018 Node* progress = StoreNode::Ideal(phase, can_reshape); 3019 if (progress != nullptr) return progress; 3020 3021 Node* my_store = in(MemNode::OopStore); 3022 if (my_store->is_MergeMem()) { 3023 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 3024 set_req_X(MemNode::OopStore, mem, phase); 3025 return this; 3026 } 3027 3028 return nullptr; 3029 } 3030 3031 //------------------------------Value----------------------------------------- 3032 const Type* StoreCMNode::Value(PhaseGVN* phase) const { 3033 // Either input is TOP ==> the result is TOP (checked in StoreNode::Value). 3034 // If extra input is TOP ==> the result is TOP 3035 const Type* t = phase->type(in(MemNode::OopStore)); 3036 if (t == Type::TOP) { 3037 return Type::TOP; 3038 } 3039 return StoreNode::Value(phase); 3040 } 3041 3042 3043 //============================================================================= 3044 //----------------------------------SCMemProjNode------------------------------ 3045 const Type* SCMemProjNode::Value(PhaseGVN* phase) const 3046 { 3047 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) { 3048 return Type::TOP; 3049 } 3050 return bottom_type(); 3051 } 3052 3053 //============================================================================= 3054 //----------------------------------LoadStoreNode------------------------------ 3055 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 3056 : Node(required), 3057 _type(rt), 3058 _adr_type(at), 3059 _barrier_data(0) 3060 { 3061 init_req(MemNode::Control, c ); 3062 init_req(MemNode::Memory , mem); 3063 init_req(MemNode::Address, adr); 3064 init_req(MemNode::ValueIn, val); 3065 init_class_id(Class_LoadStore); 3066 } 3067 3068 //------------------------------Value----------------------------------------- 3069 const Type* LoadStoreNode::Value(PhaseGVN* phase) const { 3070 // Either input is TOP ==> the result is TOP 3071 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) { 3072 return Type::TOP; 3073 } 3074 const Type* t = phase->type(in(MemNode::Memory)); 3075 if (t == Type::TOP) { 3076 return Type::TOP; 3077 } 3078 t = phase->type(in(MemNode::Address)); 3079 if (t == Type::TOP) { 3080 return Type::TOP; 3081 } 3082 t = phase->type(in(MemNode::ValueIn)); 3083 if (t == Type::TOP) { 3084 return Type::TOP; 3085 } 3086 return bottom_type(); 3087 } 3088 3089 uint LoadStoreNode::ideal_reg() const { 3090 return _type->ideal_reg(); 3091 } 3092 3093 // This method conservatively checks if the result of a LoadStoreNode is 3094 // used, that is, if it returns true, then it is definitely the case that 3095 // the result of the node is not needed. 3096 // For example, GetAndAdd can be matched into a lock_add instead of a 3097 // lock_xadd if the result of LoadStoreNode::result_not_used() is true 3098 bool LoadStoreNode::result_not_used() const { 3099 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 3100 Node *x = fast_out(i); 3101 if (x->Opcode() == Op_SCMemProj) { 3102 continue; 3103 } 3104 if (x->bottom_type() == TypeTuple::MEMBAR && 3105 !x->is_Call() && 3106 x->Opcode() != Op_Blackhole) { 3107 continue; 3108 } 3109 return false; 3110 } 3111 return true; 3112 } 3113 3114 MemBarNode* LoadStoreNode::trailing_membar() const { 3115 MemBarNode* trailing = nullptr; 3116 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 3117 Node* u = fast_out(i); 3118 if (u->is_MemBar()) { 3119 if (u->as_MemBar()->trailing_load_store()) { 3120 assert(u->Opcode() == Op_MemBarAcquire, ""); 3121 assert(trailing == nullptr, "only one"); 3122 trailing = u->as_MemBar(); 3123 #ifdef ASSERT 3124 Node* leading = trailing->leading_membar(); 3125 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar"); 3126 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair"); 3127 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair"); 3128 #endif 3129 } else { 3130 assert(u->as_MemBar()->standalone(), "wrong barrier kind"); 3131 } 3132 } 3133 } 3134 3135 return trailing; 3136 } 3137 3138 uint LoadStoreNode::size_of() const { return sizeof(*this); } 3139 3140 //============================================================================= 3141 //----------------------------------LoadStoreConditionalNode-------------------- 3142 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) { 3143 init_req(ExpectedIn, ex ); 3144 } 3145 3146 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const { 3147 // Either input is TOP ==> the result is TOP 3148 const Type* t = phase->type(in(ExpectedIn)); 3149 if (t == Type::TOP) { 3150 return Type::TOP; 3151 } 3152 return LoadStoreNode::Value(phase); 3153 } 3154 3155 //============================================================================= 3156 //-------------------------------adr_type-------------------------------------- 3157 const TypePtr* ClearArrayNode::adr_type() const { 3158 Node *adr = in(3); 3159 if (adr == nullptr) return nullptr; // node is dead 3160 return MemNode::calculate_adr_type(adr->bottom_type()); 3161 } 3162 3163 //------------------------------match_edge------------------------------------- 3164 // Do we Match on this edge index or not? Do not match memory 3165 uint ClearArrayNode::match_edge(uint idx) const { 3166 return idx > 1; 3167 } 3168 3169 //------------------------------Identity--------------------------------------- 3170 // Clearing a zero length array does nothing 3171 Node* ClearArrayNode::Identity(PhaseGVN* phase) { 3172 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 3173 } 3174 3175 //------------------------------Idealize--------------------------------------- 3176 // Clearing a short array is faster with stores 3177 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3178 // Already know this is a large node, do not try to ideal it 3179 if (_is_large) return nullptr; 3180 3181 const int unit = BytesPerLong; 3182 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 3183 if (!t) return nullptr; 3184 if (!t->is_con()) return nullptr; 3185 intptr_t raw_count = t->get_con(); 3186 intptr_t size = raw_count; 3187 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 3188 // Clearing nothing uses the Identity call. 3189 // Negative clears are possible on dead ClearArrays 3190 // (see jck test stmt114.stmt11402.val). 3191 if (size <= 0 || size % unit != 0) return nullptr; 3192 intptr_t count = size / unit; 3193 // Length too long; communicate this to matchers and assemblers. 3194 // Assemblers are responsible to produce fast hardware clears for it. 3195 if (size > InitArrayShortSize) { 3196 return new ClearArrayNode(in(0), in(1), in(2), in(3), true); 3197 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) { 3198 return nullptr; 3199 } 3200 if (!IdealizeClearArrayNode) return nullptr; 3201 Node *mem = in(1); 3202 if( phase->type(mem)==Type::TOP ) return nullptr; 3203 Node *adr = in(3); 3204 const Type* at = phase->type(adr); 3205 if( at==Type::TOP ) return nullptr; 3206 const TypePtr* atp = at->isa_ptr(); 3207 // adjust atp to be the correct array element address type 3208 if (atp == nullptr) atp = TypePtr::BOTTOM; 3209 else atp = atp->add_offset(Type::OffsetBot); 3210 // Get base for derived pointer purposes 3211 if( adr->Opcode() != Op_AddP ) Unimplemented(); 3212 Node *base = adr->in(1); 3213 3214 Node *zero = phase->makecon(TypeLong::ZERO); 3215 Node *off = phase->MakeConX(BytesPerLong); 3216 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 3217 count--; 3218 while( count-- ) { 3219 mem = phase->transform(mem); 3220 adr = phase->transform(new AddPNode(base,adr,off)); 3221 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 3222 } 3223 return mem; 3224 } 3225 3226 //----------------------------step_through---------------------------------- 3227 // Return allocation input memory edge if it is different instance 3228 // or itself if it is the one we are looking for. 3229 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) { 3230 Node* n = *np; 3231 assert(n->is_ClearArray(), "sanity"); 3232 intptr_t offset; 3233 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 3234 // This method is called only before Allocate nodes are expanded 3235 // during macro nodes expansion. Before that ClearArray nodes are 3236 // only generated in PhaseMacroExpand::generate_arraycopy() (before 3237 // Allocate nodes are expanded) which follows allocations. 3238 assert(alloc != nullptr, "should have allocation"); 3239 if (alloc->_idx == instance_id) { 3240 // Can not bypass initialization of the instance we are looking for. 3241 return false; 3242 } 3243 // Otherwise skip it. 3244 InitializeNode* init = alloc->initialization(); 3245 if (init != nullptr) 3246 *np = init->in(TypeFunc::Memory); 3247 else 3248 *np = alloc->in(TypeFunc::Memory); 3249 return true; 3250 } 3251 3252 //----------------------------clear_memory------------------------------------- 3253 // Generate code to initialize object storage to zero. 3254 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3255 intptr_t start_offset, 3256 Node* end_offset, 3257 PhaseGVN* phase) { 3258 intptr_t offset = start_offset; 3259 3260 int unit = BytesPerLong; 3261 if ((offset % unit) != 0) { 3262 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 3263 adr = phase->transform(adr); 3264 const TypePtr* atp = TypeRawPtr::BOTTOM; 3265 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3266 mem = phase->transform(mem); 3267 offset += BytesPerInt; 3268 } 3269 assert((offset % unit) == 0, ""); 3270 3271 // Initialize the remaining stuff, if any, with a ClearArray. 3272 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 3273 } 3274 3275 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3276 Node* start_offset, 3277 Node* end_offset, 3278 PhaseGVN* phase) { 3279 if (start_offset == end_offset) { 3280 // nothing to do 3281 return mem; 3282 } 3283 3284 int unit = BytesPerLong; 3285 Node* zbase = start_offset; 3286 Node* zend = end_offset; 3287 3288 // Scale to the unit required by the CPU: 3289 if (!Matcher::init_array_count_is_in_bytes) { 3290 Node* shift = phase->intcon(exact_log2(unit)); 3291 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 3292 zend = phase->transform(new URShiftXNode(zend, shift) ); 3293 } 3294 3295 // Bulk clear double-words 3296 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 3297 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 3298 mem = new ClearArrayNode(ctl, mem, zsize, adr, false); 3299 return phase->transform(mem); 3300 } 3301 3302 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 3303 intptr_t start_offset, 3304 intptr_t end_offset, 3305 PhaseGVN* phase) { 3306 if (start_offset == end_offset) { 3307 // nothing to do 3308 return mem; 3309 } 3310 3311 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 3312 intptr_t done_offset = end_offset; 3313 if ((done_offset % BytesPerLong) != 0) { 3314 done_offset -= BytesPerInt; 3315 } 3316 if (done_offset > start_offset) { 3317 mem = clear_memory(ctl, mem, dest, 3318 start_offset, phase->MakeConX(done_offset), phase); 3319 } 3320 if (done_offset < end_offset) { // emit the final 32-bit store 3321 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 3322 adr = phase->transform(adr); 3323 const TypePtr* atp = TypeRawPtr::BOTTOM; 3324 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 3325 mem = phase->transform(mem); 3326 done_offset += BytesPerInt; 3327 } 3328 assert(done_offset == end_offset, ""); 3329 return mem; 3330 } 3331 3332 //============================================================================= 3333 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 3334 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)), 3335 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone) 3336 #ifdef ASSERT 3337 , _pair_idx(0) 3338 #endif 3339 { 3340 init_class_id(Class_MemBar); 3341 Node* top = C->top(); 3342 init_req(TypeFunc::I_O,top); 3343 init_req(TypeFunc::FramePtr,top); 3344 init_req(TypeFunc::ReturnAdr,top); 3345 if (precedent != nullptr) 3346 init_req(TypeFunc::Parms, precedent); 3347 } 3348 3349 //------------------------------cmp-------------------------------------------- 3350 uint MemBarNode::hash() const { return NO_HASH; } 3351 bool MemBarNode::cmp( const Node &n ) const { 3352 return (&n == this); // Always fail except on self 3353 } 3354 3355 //------------------------------make------------------------------------------- 3356 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 3357 switch (opcode) { 3358 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 3359 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 3360 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 3361 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 3362 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 3363 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn); 3364 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 3365 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 3366 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 3367 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 3368 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn); 3369 case Op_Initialize: return new InitializeNode(C, atp, pn); 3370 default: ShouldNotReachHere(); return nullptr; 3371 } 3372 } 3373 3374 void MemBarNode::remove(PhaseIterGVN *igvn) { 3375 if (outcnt() != 2) { 3376 assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory"); 3377 assert(outcnt() == 1, "Only control then"); 3378 } 3379 if (trailing_store() || trailing_load_store()) { 3380 MemBarNode* leading = leading_membar(); 3381 if (leading != nullptr) { 3382 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars"); 3383 leading->remove(igvn); 3384 } 3385 } 3386 if (proj_out_or_null(TypeFunc::Memory) != nullptr) { 3387 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 3388 } 3389 if (proj_out_or_null(TypeFunc::Control) != nullptr) { 3390 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 3391 } 3392 } 3393 3394 //------------------------------Ideal------------------------------------------ 3395 // Return a node which is more "ideal" than the current node. Strip out 3396 // control copies 3397 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 3398 if (remove_dead_region(phase, can_reshape)) return this; 3399 // Don't bother trying to transform a dead node 3400 if (in(0) && in(0)->is_top()) { 3401 return nullptr; 3402 } 3403 3404 bool progress = false; 3405 // Eliminate volatile MemBars for scalar replaced objects. 3406 if (can_reshape && req() == (Precedent+1)) { 3407 bool eliminate = false; 3408 int opc = Opcode(); 3409 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 3410 // Volatile field loads and stores. 3411 Node* my_mem = in(MemBarNode::Precedent); 3412 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 3413 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 3414 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 3415 // replace this Precedent (decodeN) with the Load instead. 3416 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 3417 Node* load_node = my_mem->in(1); 3418 set_req(MemBarNode::Precedent, load_node); 3419 phase->is_IterGVN()->_worklist.push(my_mem); 3420 my_mem = load_node; 3421 } else { 3422 assert(my_mem->unique_out() == this, "sanity"); 3423 assert(!trailing_load_store(), "load store node can't be eliminated"); 3424 del_req(Precedent); 3425 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 3426 my_mem = nullptr; 3427 } 3428 progress = true; 3429 } 3430 if (my_mem != nullptr && my_mem->is_Mem()) { 3431 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 3432 // Check for scalar replaced object reference. 3433 if( t_oop != nullptr && t_oop->is_known_instance_field() && 3434 t_oop->offset() != Type::OffsetBot && 3435 t_oop->offset() != Type::OffsetTop) { 3436 eliminate = true; 3437 } 3438 } 3439 } else if (opc == Op_MemBarRelease) { 3440 // Final field stores. 3441 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent)); 3442 if ((alloc != nullptr) && alloc->is_Allocate() && 3443 alloc->as_Allocate()->does_not_escape_thread()) { 3444 // The allocated object does not escape. 3445 eliminate = true; 3446 } 3447 } 3448 if (eliminate) { 3449 // Replace MemBar projections by its inputs. 3450 PhaseIterGVN* igvn = phase->is_IterGVN(); 3451 remove(igvn); 3452 // Must return either the original node (now dead) or a new node 3453 // (Do not return a top here, since that would break the uniqueness of top.) 3454 return new ConINode(TypeInt::ZERO); 3455 } 3456 } 3457 return progress ? this : nullptr; 3458 } 3459 3460 //------------------------------Value------------------------------------------ 3461 const Type* MemBarNode::Value(PhaseGVN* phase) const { 3462 if( !in(0) ) return Type::TOP; 3463 if( phase->type(in(0)) == Type::TOP ) 3464 return Type::TOP; 3465 return TypeTuple::MEMBAR; 3466 } 3467 3468 //------------------------------match------------------------------------------ 3469 // Construct projections for memory. 3470 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 3471 switch (proj->_con) { 3472 case TypeFunc::Control: 3473 case TypeFunc::Memory: 3474 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3475 } 3476 ShouldNotReachHere(); 3477 return nullptr; 3478 } 3479 3480 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3481 trailing->_kind = TrailingStore; 3482 leading->_kind = LeadingStore; 3483 #ifdef ASSERT 3484 trailing->_pair_idx = leading->_idx; 3485 leading->_pair_idx = leading->_idx; 3486 #endif 3487 } 3488 3489 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) { 3490 trailing->_kind = TrailingLoadStore; 3491 leading->_kind = LeadingLoadStore; 3492 #ifdef ASSERT 3493 trailing->_pair_idx = leading->_idx; 3494 leading->_pair_idx = leading->_idx; 3495 #endif 3496 } 3497 3498 MemBarNode* MemBarNode::trailing_membar() const { 3499 ResourceMark rm; 3500 Node* trailing = (Node*)this; 3501 VectorSet seen; 3502 Node_Stack multis(0); 3503 do { 3504 Node* c = trailing; 3505 uint i = 0; 3506 do { 3507 trailing = nullptr; 3508 for (; i < c->outcnt(); i++) { 3509 Node* next = c->raw_out(i); 3510 if (next != c && next->is_CFG()) { 3511 if (c->is_MultiBranch()) { 3512 if (multis.node() == c) { 3513 multis.set_index(i+1); 3514 } else { 3515 multis.push(c, i+1); 3516 } 3517 } 3518 trailing = next; 3519 break; 3520 } 3521 } 3522 if (trailing != nullptr && !seen.test_set(trailing->_idx)) { 3523 break; 3524 } 3525 while (multis.size() > 0) { 3526 c = multis.node(); 3527 i = multis.index(); 3528 if (i < c->req()) { 3529 break; 3530 } 3531 multis.pop(); 3532 } 3533 } while (multis.size() > 0); 3534 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing()); 3535 3536 MemBarNode* mb = trailing->as_MemBar(); 3537 assert((mb->_kind == TrailingStore && _kind == LeadingStore) || 3538 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar"); 3539 assert(mb->_pair_idx == _pair_idx, "bad trailing membar"); 3540 return mb; 3541 } 3542 3543 MemBarNode* MemBarNode::leading_membar() const { 3544 ResourceMark rm; 3545 VectorSet seen; 3546 Node_Stack regions(0); 3547 Node* leading = in(0); 3548 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) { 3549 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) { 3550 leading = nullptr; 3551 while (regions.size() > 0 && leading == nullptr) { 3552 Node* r = regions.node(); 3553 uint i = regions.index(); 3554 if (i < r->req()) { 3555 leading = r->in(i); 3556 regions.set_index(i+1); 3557 } else { 3558 regions.pop(); 3559 } 3560 } 3561 if (leading == nullptr) { 3562 assert(regions.size() == 0, "all paths should have been tried"); 3563 return nullptr; 3564 } 3565 } 3566 if (leading->is_Region()) { 3567 regions.push(leading, 2); 3568 leading = leading->in(1); 3569 } else { 3570 leading = leading->in(0); 3571 } 3572 } 3573 #ifdef ASSERT 3574 Unique_Node_List wq; 3575 wq.push((Node*)this); 3576 uint found = 0; 3577 for (uint i = 0; i < wq.size(); i++) { 3578 Node* n = wq.at(i); 3579 if (n->is_Region()) { 3580 for (uint j = 1; j < n->req(); j++) { 3581 Node* in = n->in(j); 3582 if (in != nullptr && !in->is_top()) { 3583 wq.push(in); 3584 } 3585 } 3586 } else { 3587 if (n->is_MemBar() && n->as_MemBar()->leading()) { 3588 assert(n == leading, "consistency check failed"); 3589 found++; 3590 } else { 3591 Node* in = n->in(0); 3592 if (in != nullptr && !in->is_top()) { 3593 wq.push(in); 3594 } 3595 } 3596 } 3597 } 3598 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed"); 3599 #endif 3600 if (leading == nullptr) { 3601 return nullptr; 3602 } 3603 MemBarNode* mb = leading->as_MemBar(); 3604 assert((mb->_kind == LeadingStore && _kind == TrailingStore) || 3605 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar"); 3606 assert(mb->_pair_idx == _pair_idx, "bad leading membar"); 3607 return mb; 3608 } 3609 3610 3611 //===========================InitializeNode==================================== 3612 // SUMMARY: 3613 // This node acts as a memory barrier on raw memory, after some raw stores. 3614 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3615 // The Initialize can 'capture' suitably constrained stores as raw inits. 3616 // It can coalesce related raw stores into larger units (called 'tiles'). 3617 // It can avoid zeroing new storage for memory units which have raw inits. 3618 // At macro-expansion, it is marked 'complete', and does not optimize further. 3619 // 3620 // EXAMPLE: 3621 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3622 // ctl = incoming control; mem* = incoming memory 3623 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3624 // First allocate uninitialized memory and fill in the header: 3625 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3626 // ctl := alloc.Control; mem* := alloc.Memory* 3627 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3628 // Then initialize to zero the non-header parts of the raw memory block: 3629 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3630 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3631 // After the initialize node executes, the object is ready for service: 3632 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3633 // Suppose its body is immediately initialized as {1,2}: 3634 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3635 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3636 // mem.SLICE(#short[*]) := store2 3637 // 3638 // DETAILS: 3639 // An InitializeNode collects and isolates object initialization after 3640 // an AllocateNode and before the next possible safepoint. As a 3641 // memory barrier (MemBarNode), it keeps critical stores from drifting 3642 // down past any safepoint or any publication of the allocation. 3643 // Before this barrier, a newly-allocated object may have uninitialized bits. 3644 // After this barrier, it may be treated as a real oop, and GC is allowed. 3645 // 3646 // The semantics of the InitializeNode include an implicit zeroing of 3647 // the new object from object header to the end of the object. 3648 // (The object header and end are determined by the AllocateNode.) 3649 // 3650 // Certain stores may be added as direct inputs to the InitializeNode. 3651 // These stores must update raw memory, and they must be to addresses 3652 // derived from the raw address produced by AllocateNode, and with 3653 // a constant offset. They must be ordered by increasing offset. 3654 // The first one is at in(RawStores), the last at in(req()-1). 3655 // Unlike most memory operations, they are not linked in a chain, 3656 // but are displayed in parallel as users of the rawmem output of 3657 // the allocation. 3658 // 3659 // (See comments in InitializeNode::capture_store, which continue 3660 // the example given above.) 3661 // 3662 // When the associated Allocate is macro-expanded, the InitializeNode 3663 // may be rewritten to optimize collected stores. A ClearArrayNode 3664 // may also be created at that point to represent any required zeroing. 3665 // The InitializeNode is then marked 'complete', prohibiting further 3666 // capturing of nearby memory operations. 3667 // 3668 // During macro-expansion, all captured initializations which store 3669 // constant values of 32 bits or smaller are coalesced (if advantageous) 3670 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3671 // initialized in fewer memory operations. Memory words which are 3672 // covered by neither tiles nor non-constant stores are pre-zeroed 3673 // by explicit stores of zero. (The code shape happens to do all 3674 // zeroing first, then all other stores, with both sequences occurring 3675 // in order of ascending offsets.) 3676 // 3677 // Alternatively, code may be inserted between an AllocateNode and its 3678 // InitializeNode, to perform arbitrary initialization of the new object. 3679 // E.g., the object copying intrinsics insert complex data transfers here. 3680 // The initialization must then be marked as 'complete' disable the 3681 // built-in zeroing semantics and the collection of initializing stores. 3682 // 3683 // While an InitializeNode is incomplete, reads from the memory state 3684 // produced by it are optimizable if they match the control edge and 3685 // new oop address associated with the allocation/initialization. 3686 // They return a stored value (if the offset matches) or else zero. 3687 // A write to the memory state, if it matches control and address, 3688 // and if it is to a constant offset, may be 'captured' by the 3689 // InitializeNode. It is cloned as a raw memory operation and rewired 3690 // inside the initialization, to the raw oop produced by the allocation. 3691 // Operations on addresses which are provably distinct (e.g., to 3692 // other AllocateNodes) are allowed to bypass the initialization. 3693 // 3694 // The effect of all this is to consolidate object initialization 3695 // (both arrays and non-arrays, both piecewise and bulk) into a 3696 // single location, where it can be optimized as a unit. 3697 // 3698 // Only stores with an offset less than TrackedInitializationLimit words 3699 // will be considered for capture by an InitializeNode. This puts a 3700 // reasonable limit on the complexity of optimized initializations. 3701 3702 //---------------------------InitializeNode------------------------------------ 3703 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3704 : MemBarNode(C, adr_type, rawoop), 3705 _is_complete(Incomplete), _does_not_escape(false) 3706 { 3707 init_class_id(Class_Initialize); 3708 3709 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3710 assert(in(RawAddress) == rawoop, "proper init"); 3711 // Note: allocation() can be null, for secondary initialization barriers 3712 } 3713 3714 // Since this node is not matched, it will be processed by the 3715 // register allocator. Declare that there are no constraints 3716 // on the allocation of the RawAddress edge. 3717 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3718 // This edge should be set to top, by the set_complete. But be conservative. 3719 if (idx == InitializeNode::RawAddress) 3720 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3721 return RegMask::Empty; 3722 } 3723 3724 Node* InitializeNode::memory(uint alias_idx) { 3725 Node* mem = in(Memory); 3726 if (mem->is_MergeMem()) { 3727 return mem->as_MergeMem()->memory_at(alias_idx); 3728 } else { 3729 // incoming raw memory is not split 3730 return mem; 3731 } 3732 } 3733 3734 bool InitializeNode::is_non_zero() { 3735 if (is_complete()) return false; 3736 remove_extra_zeroes(); 3737 return (req() > RawStores); 3738 } 3739 3740 void InitializeNode::set_complete(PhaseGVN* phase) { 3741 assert(!is_complete(), "caller responsibility"); 3742 _is_complete = Complete; 3743 3744 // After this node is complete, it contains a bunch of 3745 // raw-memory initializations. There is no need for 3746 // it to have anything to do with non-raw memory effects. 3747 // Therefore, tell all non-raw users to re-optimize themselves, 3748 // after skipping the memory effects of this initialization. 3749 PhaseIterGVN* igvn = phase->is_IterGVN(); 3750 if (igvn) igvn->add_users_to_worklist(this); 3751 } 3752 3753 // convenience function 3754 // return false if the init contains any stores already 3755 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3756 InitializeNode* init = initialization(); 3757 if (init == nullptr || init->is_complete()) return false; 3758 init->remove_extra_zeroes(); 3759 // for now, if this allocation has already collected any inits, bail: 3760 if (init->is_non_zero()) return false; 3761 init->set_complete(phase); 3762 return true; 3763 } 3764 3765 void InitializeNode::remove_extra_zeroes() { 3766 if (req() == RawStores) return; 3767 Node* zmem = zero_memory(); 3768 uint fill = RawStores; 3769 for (uint i = fill; i < req(); i++) { 3770 Node* n = in(i); 3771 if (n->is_top() || n == zmem) continue; // skip 3772 if (fill < i) set_req(fill, n); // compact 3773 ++fill; 3774 } 3775 // delete any empty spaces created: 3776 while (fill < req()) { 3777 del_req(fill); 3778 } 3779 } 3780 3781 // Helper for remembering which stores go with which offsets. 3782 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) { 3783 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3784 intptr_t offset = -1; 3785 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3786 phase, offset); 3787 if (base == nullptr) return -1; // something is dead, 3788 if (offset < 0) return -1; // dead, dead 3789 return offset; 3790 } 3791 3792 // Helper for proving that an initialization expression is 3793 // "simple enough" to be folded into an object initialization. 3794 // Attempts to prove that a store's initial value 'n' can be captured 3795 // within the initialization without creating a vicious cycle, such as: 3796 // { Foo p = new Foo(); p.next = p; } 3797 // True for constants and parameters and small combinations thereof. 3798 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) { 3799 ResourceMark rm; 3800 Unique_Node_List worklist; 3801 worklist.push(value); 3802 3803 uint complexity_limit = 20; 3804 for (uint j = 0; j < worklist.size(); j++) { 3805 if (j >= complexity_limit) { 3806 return false; // Bail out if processed too many nodes 3807 } 3808 3809 Node* n = worklist.at(j); 3810 if (n == nullptr) continue; // (can this really happen?) 3811 if (n->is_Proj()) n = n->in(0); 3812 if (n == this) return false; // found a cycle 3813 if (n->is_Con()) continue; 3814 if (n->is_Start()) continue; // params, etc., are OK 3815 if (n->is_Root()) continue; // even better 3816 3817 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation 3818 if (n->is_CFG() && phase->is_dominator(n, allocation())) { 3819 continue; 3820 } 3821 3822 Node* ctl = n->in(0); 3823 if (ctl != nullptr && !ctl->is_top()) { 3824 if (ctl->is_Proj()) ctl = ctl->in(0); 3825 if (ctl == this) return false; 3826 3827 // If we already know that the enclosing memory op is pinned right after 3828 // the init, then any control flow that the store has picked up 3829 // must have preceded the init, or else be equal to the init. 3830 // Even after loop optimizations (which might change control edges) 3831 // a store is never pinned *before* the availability of its inputs. 3832 if (!MemNode::all_controls_dominate(n, this)) 3833 return false; // failed to prove a good control 3834 } 3835 3836 // Check data edges for possible dependencies on 'this'. 3837 for (uint i = 1; i < n->req(); i++) { 3838 Node* m = n->in(i); 3839 if (m == nullptr || m == n || m->is_top()) continue; 3840 3841 // Only process data inputs once 3842 worklist.push(m); 3843 } 3844 } 3845 3846 return true; 3847 } 3848 3849 // Here are all the checks a Store must pass before it can be moved into 3850 // an initialization. Returns zero if a check fails. 3851 // On success, returns the (constant) offset to which the store applies, 3852 // within the initialized memory. 3853 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) { 3854 const int FAIL = 0; 3855 if (st->req() != MemNode::ValueIn + 1) 3856 return FAIL; // an inscrutable StoreNode (card mark?) 3857 Node* ctl = st->in(MemNode::Control); 3858 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this)) 3859 return FAIL; // must be unconditional after the initialization 3860 Node* mem = st->in(MemNode::Memory); 3861 if (!(mem->is_Proj() && mem->in(0) == this)) 3862 return FAIL; // must not be preceded by other stores 3863 Node* adr = st->in(MemNode::Address); 3864 intptr_t offset; 3865 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3866 if (alloc == nullptr) 3867 return FAIL; // inscrutable address 3868 if (alloc != allocation()) 3869 return FAIL; // wrong allocation! (store needs to float up) 3870 int size_in_bytes = st->memory_size(); 3871 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) { 3872 return FAIL; // mismatched access 3873 } 3874 Node* val = st->in(MemNode::ValueIn); 3875 3876 if (!detect_init_independence(val, phase)) 3877 return FAIL; // stored value must be 'simple enough' 3878 3879 // The Store can be captured only if nothing after the allocation 3880 // and before the Store is using the memory location that the store 3881 // overwrites. 3882 bool failed = false; 3883 // If is_complete_with_arraycopy() is true the shape of the graph is 3884 // well defined and is safe so no need for extra checks. 3885 if (!is_complete_with_arraycopy()) { 3886 // We are going to look at each use of the memory state following 3887 // the allocation to make sure nothing reads the memory that the 3888 // Store writes. 3889 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3890 int alias_idx = phase->C->get_alias_index(t_adr); 3891 ResourceMark rm; 3892 Unique_Node_List mems; 3893 mems.push(mem); 3894 Node* unique_merge = nullptr; 3895 for (uint next = 0; next < mems.size(); ++next) { 3896 Node *m = mems.at(next); 3897 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3898 Node *n = m->fast_out(j); 3899 if (n->outcnt() == 0) { 3900 continue; 3901 } 3902 if (n == st) { 3903 continue; 3904 } else if (n->in(0) != nullptr && n->in(0) != ctl) { 3905 // If the control of this use is different from the control 3906 // of the Store which is right after the InitializeNode then 3907 // this node cannot be between the InitializeNode and the 3908 // Store. 3909 continue; 3910 } else if (n->is_MergeMem()) { 3911 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3912 // We can hit a MergeMemNode (that will likely go away 3913 // later) that is a direct use of the memory state 3914 // following the InitializeNode on the same slice as the 3915 // store node that we'd like to capture. We need to check 3916 // the uses of the MergeMemNode. 3917 mems.push(n); 3918 } 3919 } else if (n->is_Mem()) { 3920 Node* other_adr = n->in(MemNode::Address); 3921 if (other_adr == adr) { 3922 failed = true; 3923 break; 3924 } else { 3925 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3926 if (other_t_adr != nullptr) { 3927 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3928 if (other_alias_idx == alias_idx) { 3929 // A load from the same memory slice as the store right 3930 // after the InitializeNode. We check the control of the 3931 // object/array that is loaded from. If it's the same as 3932 // the store control then we cannot capture the store. 3933 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3934 Node* base = other_adr; 3935 assert(base->is_AddP(), "should be addp but is %s", base->Name()); 3936 base = base->in(AddPNode::Base); 3937 if (base != nullptr) { 3938 base = base->uncast(); 3939 if (base->is_Proj() && base->in(0) == alloc) { 3940 failed = true; 3941 break; 3942 } 3943 } 3944 } 3945 } 3946 } 3947 } else { 3948 failed = true; 3949 break; 3950 } 3951 } 3952 } 3953 } 3954 if (failed) { 3955 if (!can_reshape) { 3956 // We decided we couldn't capture the store during parsing. We 3957 // should try again during the next IGVN once the graph is 3958 // cleaner. 3959 phase->C->record_for_igvn(st); 3960 } 3961 return FAIL; 3962 } 3963 3964 return offset; // success 3965 } 3966 3967 // Find the captured store in(i) which corresponds to the range 3968 // [start..start+size) in the initialized object. 3969 // If there is one, return its index i. If there isn't, return the 3970 // negative of the index where it should be inserted. 3971 // Return 0 if the queried range overlaps an initialization boundary 3972 // or if dead code is encountered. 3973 // If size_in_bytes is zero, do not bother with overlap checks. 3974 int InitializeNode::captured_store_insertion_point(intptr_t start, 3975 int size_in_bytes, 3976 PhaseValues* phase) { 3977 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize); 3978 3979 if (is_complete()) 3980 return FAIL; // arraycopy got here first; punt 3981 3982 assert(allocation() != nullptr, "must be present"); 3983 3984 // no negatives, no header fields: 3985 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3986 3987 // after a certain size, we bail out on tracking all the stores: 3988 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3989 if (start >= ti_limit) return FAIL; 3990 3991 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3992 if (i >= limit) return -(int)i; // not found; here is where to put it 3993 3994 Node* st = in(i); 3995 intptr_t st_off = get_store_offset(st, phase); 3996 if (st_off < 0) { 3997 if (st != zero_memory()) { 3998 return FAIL; // bail out if there is dead garbage 3999 } 4000 } else if (st_off > start) { 4001 // ...we are done, since stores are ordered 4002 if (st_off < start + size_in_bytes) { 4003 return FAIL; // the next store overlaps 4004 } 4005 return -(int)i; // not found; here is where to put it 4006 } else if (st_off < start) { 4007 assert(st->as_Store()->memory_size() <= MAX_STORE, ""); 4008 if (size_in_bytes != 0 && 4009 start < st_off + MAX_STORE && 4010 start < st_off + st->as_Store()->memory_size()) { 4011 return FAIL; // the previous store overlaps 4012 } 4013 } else { 4014 if (size_in_bytes != 0 && 4015 st->as_Store()->memory_size() != size_in_bytes) { 4016 return FAIL; // mismatched store size 4017 } 4018 return i; 4019 } 4020 4021 ++i; 4022 } 4023 } 4024 4025 // Look for a captured store which initializes at the offset 'start' 4026 // with the given size. If there is no such store, and no other 4027 // initialization interferes, then return zero_memory (the memory 4028 // projection of the AllocateNode). 4029 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 4030 PhaseValues* phase) { 4031 assert(stores_are_sane(phase), ""); 4032 int i = captured_store_insertion_point(start, size_in_bytes, phase); 4033 if (i == 0) { 4034 return nullptr; // something is dead 4035 } else if (i < 0) { 4036 return zero_memory(); // just primordial zero bits here 4037 } else { 4038 Node* st = in(i); // here is the store at this position 4039 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 4040 return st; 4041 } 4042 } 4043 4044 // Create, as a raw pointer, an address within my new object at 'offset'. 4045 Node* InitializeNode::make_raw_address(intptr_t offset, 4046 PhaseGVN* phase) { 4047 Node* addr = in(RawAddress); 4048 if (offset != 0) { 4049 Compile* C = phase->C; 4050 addr = phase->transform( new AddPNode(C->top(), addr, 4051 phase->MakeConX(offset)) ); 4052 } 4053 return addr; 4054 } 4055 4056 // Clone the given store, converting it into a raw store 4057 // initializing a field or element of my new object. 4058 // Caller is responsible for retiring the original store, 4059 // with subsume_node or the like. 4060 // 4061 // From the example above InitializeNode::InitializeNode, 4062 // here are the old stores to be captured: 4063 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 4064 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 4065 // 4066 // Here is the changed code; note the extra edges on init: 4067 // alloc = (Allocate ...) 4068 // rawoop = alloc.RawAddress 4069 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 4070 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 4071 // init = (Initialize alloc.Control alloc.Memory rawoop 4072 // rawstore1 rawstore2) 4073 // 4074 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 4075 PhaseGVN* phase, bool can_reshape) { 4076 assert(stores_are_sane(phase), ""); 4077 4078 if (start < 0) return nullptr; 4079 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 4080 4081 Compile* C = phase->C; 4082 int size_in_bytes = st->memory_size(); 4083 int i = captured_store_insertion_point(start, size_in_bytes, phase); 4084 if (i == 0) return nullptr; // bail out 4085 Node* prev_mem = nullptr; // raw memory for the captured store 4086 if (i > 0) { 4087 prev_mem = in(i); // there is a pre-existing store under this one 4088 set_req(i, C->top()); // temporarily disconnect it 4089 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 4090 } else { 4091 i = -i; // no pre-existing store 4092 prev_mem = zero_memory(); // a slice of the newly allocated object 4093 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 4094 set_req(--i, C->top()); // reuse this edge; it has been folded away 4095 else 4096 ins_req(i, C->top()); // build a new edge 4097 } 4098 Node* new_st = st->clone(); 4099 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4100 new_st->set_req(MemNode::Control, in(Control)); 4101 new_st->set_req(MemNode::Memory, prev_mem); 4102 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 4103 bs->eliminate_gc_barrier_data(new_st); 4104 new_st = phase->transform(new_st); 4105 4106 // At this point, new_st might have swallowed a pre-existing store 4107 // at the same offset, or perhaps new_st might have disappeared, 4108 // if it redundantly stored the same value (or zero to fresh memory). 4109 4110 // In any case, wire it in: 4111 PhaseIterGVN* igvn = phase->is_IterGVN(); 4112 if (igvn) { 4113 igvn->rehash_node_delayed(this); 4114 } 4115 set_req(i, new_st); 4116 4117 // The caller may now kill the old guy. 4118 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 4119 assert(check_st == new_st || check_st == nullptr, "must be findable"); 4120 assert(!is_complete(), ""); 4121 return new_st; 4122 } 4123 4124 static bool store_constant(jlong* tiles, int num_tiles, 4125 intptr_t st_off, int st_size, 4126 jlong con) { 4127 if ((st_off & (st_size-1)) != 0) 4128 return false; // strange store offset (assume size==2**N) 4129 address addr = (address)tiles + st_off; 4130 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 4131 switch (st_size) { 4132 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 4133 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 4134 case sizeof(jint): *(jint*) addr = (jint) con; break; 4135 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 4136 default: return false; // strange store size (detect size!=2**N here) 4137 } 4138 return true; // return success to caller 4139 } 4140 4141 // Coalesce subword constants into int constants and possibly 4142 // into long constants. The goal, if the CPU permits, 4143 // is to initialize the object with a small number of 64-bit tiles. 4144 // Also, convert floating-point constants to bit patterns. 4145 // Non-constants are not relevant to this pass. 4146 // 4147 // In terms of the running example on InitializeNode::InitializeNode 4148 // and InitializeNode::capture_store, here is the transformation 4149 // of rawstore1 and rawstore2 into rawstore12: 4150 // alloc = (Allocate ...) 4151 // rawoop = alloc.RawAddress 4152 // tile12 = 0x00010002 4153 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 4154 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 4155 // 4156 void 4157 InitializeNode::coalesce_subword_stores(intptr_t header_size, 4158 Node* size_in_bytes, 4159 PhaseGVN* phase) { 4160 Compile* C = phase->C; 4161 4162 assert(stores_are_sane(phase), ""); 4163 // Note: After this pass, they are not completely sane, 4164 // since there may be some overlaps. 4165 4166 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 4167 4168 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 4169 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 4170 size_limit = MIN2(size_limit, ti_limit); 4171 size_limit = align_up(size_limit, BytesPerLong); 4172 int num_tiles = size_limit / BytesPerLong; 4173 4174 // allocate space for the tile map: 4175 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 4176 jlong tiles_buf[small_len]; 4177 Node* nodes_buf[small_len]; 4178 jlong inits_buf[small_len]; 4179 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 4180 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 4181 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 4182 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 4183 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 4184 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 4185 // tiles: exact bitwise model of all primitive constants 4186 // nodes: last constant-storing node subsumed into the tiles model 4187 // inits: which bytes (in each tile) are touched by any initializations 4188 4189 //// Pass A: Fill in the tile model with any relevant stores. 4190 4191 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 4192 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 4193 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 4194 Node* zmem = zero_memory(); // initially zero memory state 4195 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4196 Node* st = in(i); 4197 intptr_t st_off = get_store_offset(st, phase); 4198 4199 // Figure out the store's offset and constant value: 4200 if (st_off < header_size) continue; //skip (ignore header) 4201 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 4202 int st_size = st->as_Store()->memory_size(); 4203 if (st_off + st_size > size_limit) break; 4204 4205 // Record which bytes are touched, whether by constant or not. 4206 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 4207 continue; // skip (strange store size) 4208 4209 const Type* val = phase->type(st->in(MemNode::ValueIn)); 4210 if (!val->singleton()) continue; //skip (non-con store) 4211 BasicType type = val->basic_type(); 4212 4213 jlong con = 0; 4214 switch (type) { 4215 case T_INT: con = val->is_int()->get_con(); break; 4216 case T_LONG: con = val->is_long()->get_con(); break; 4217 case T_FLOAT: con = jint_cast(val->getf()); break; 4218 case T_DOUBLE: con = jlong_cast(val->getd()); break; 4219 default: continue; //skip (odd store type) 4220 } 4221 4222 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 4223 st->Opcode() == Op_StoreL) { 4224 continue; // This StoreL is already optimal. 4225 } 4226 4227 // Store down the constant. 4228 store_constant(tiles, num_tiles, st_off, st_size, con); 4229 4230 intptr_t j = st_off >> LogBytesPerLong; 4231 4232 if (type == T_INT && st_size == BytesPerInt 4233 && (st_off & BytesPerInt) == BytesPerInt) { 4234 jlong lcon = tiles[j]; 4235 if (!Matcher::isSimpleConstant64(lcon) && 4236 st->Opcode() == Op_StoreI) { 4237 // This StoreI is already optimal by itself. 4238 jint* intcon = (jint*) &tiles[j]; 4239 intcon[1] = 0; // undo the store_constant() 4240 4241 // If the previous store is also optimal by itself, back up and 4242 // undo the action of the previous loop iteration... if we can. 4243 // But if we can't, just let the previous half take care of itself. 4244 st = nodes[j]; 4245 st_off -= BytesPerInt; 4246 con = intcon[0]; 4247 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) { 4248 assert(st_off >= header_size, "still ignoring header"); 4249 assert(get_store_offset(st, phase) == st_off, "must be"); 4250 assert(in(i-1) == zmem, "must be"); 4251 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 4252 assert(con == tcon->is_int()->get_con(), "must be"); 4253 // Undo the effects of the previous loop trip, which swallowed st: 4254 intcon[0] = 0; // undo store_constant() 4255 set_req(i-1, st); // undo set_req(i, zmem) 4256 nodes[j] = nullptr; // undo nodes[j] = st 4257 --old_subword; // undo ++old_subword 4258 } 4259 continue; // This StoreI is already optimal. 4260 } 4261 } 4262 4263 // This store is not needed. 4264 set_req(i, zmem); 4265 nodes[j] = st; // record for the moment 4266 if (st_size < BytesPerLong) // something has changed 4267 ++old_subword; // includes int/float, but who's counting... 4268 else ++old_long; 4269 } 4270 4271 if ((old_subword + old_long) == 0) 4272 return; // nothing more to do 4273 4274 //// Pass B: Convert any non-zero tiles into optimal constant stores. 4275 // Be sure to insert them before overlapping non-constant stores. 4276 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 4277 for (int j = 0; j < num_tiles; j++) { 4278 jlong con = tiles[j]; 4279 jlong init = inits[j]; 4280 if (con == 0) continue; 4281 jint con0, con1; // split the constant, address-wise 4282 jint init0, init1; // split the init map, address-wise 4283 { union { jlong con; jint intcon[2]; } u; 4284 u.con = con; 4285 con0 = u.intcon[0]; 4286 con1 = u.intcon[1]; 4287 u.con = init; 4288 init0 = u.intcon[0]; 4289 init1 = u.intcon[1]; 4290 } 4291 4292 Node* old = nodes[j]; 4293 assert(old != nullptr, "need the prior store"); 4294 intptr_t offset = (j * BytesPerLong); 4295 4296 bool split = !Matcher::isSimpleConstant64(con); 4297 4298 if (offset < header_size) { 4299 assert(offset + BytesPerInt >= header_size, "second int counts"); 4300 assert(*(jint*)&tiles[j] == 0, "junk in header"); 4301 split = true; // only the second word counts 4302 // Example: int a[] = { 42 ... } 4303 } else if (con0 == 0 && init0 == -1) { 4304 split = true; // first word is covered by full inits 4305 // Example: int a[] = { ... foo(), 42 ... } 4306 } else if (con1 == 0 && init1 == -1) { 4307 split = true; // second word is covered by full inits 4308 // Example: int a[] = { ... 42, foo() ... } 4309 } 4310 4311 // Here's a case where init0 is neither 0 nor -1: 4312 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 4313 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 4314 // In this case the tile is not split; it is (jlong)42. 4315 // The big tile is stored down, and then the foo() value is inserted. 4316 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 4317 4318 Node* ctl = old->in(MemNode::Control); 4319 Node* adr = make_raw_address(offset, phase); 4320 const TypePtr* atp = TypeRawPtr::BOTTOM; 4321 4322 // One or two coalesced stores to plop down. 4323 Node* st[2]; 4324 intptr_t off[2]; 4325 int nst = 0; 4326 if (!split) { 4327 ++new_long; 4328 off[nst] = offset; 4329 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4330 phase->longcon(con), T_LONG, MemNode::unordered); 4331 } else { 4332 // Omit either if it is a zero. 4333 if (con0 != 0) { 4334 ++new_int; 4335 off[nst] = offset; 4336 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4337 phase->intcon(con0), T_INT, MemNode::unordered); 4338 } 4339 if (con1 != 0) { 4340 ++new_int; 4341 offset += BytesPerInt; 4342 adr = make_raw_address(offset, phase); 4343 off[nst] = offset; 4344 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 4345 phase->intcon(con1), T_INT, MemNode::unordered); 4346 } 4347 } 4348 4349 // Insert second store first, then the first before the second. 4350 // Insert each one just before any overlapping non-constant stores. 4351 while (nst > 0) { 4352 Node* st1 = st[--nst]; 4353 C->copy_node_notes_to(st1, old); 4354 st1 = phase->transform(st1); 4355 offset = off[nst]; 4356 assert(offset >= header_size, "do not smash header"); 4357 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 4358 guarantee(ins_idx != 0, "must re-insert constant store"); 4359 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 4360 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 4361 set_req(--ins_idx, st1); 4362 else 4363 ins_req(ins_idx, st1); 4364 } 4365 } 4366 4367 if (PrintCompilation && WizardMode) 4368 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 4369 old_subword, old_long, new_int, new_long); 4370 if (C->log() != nullptr) 4371 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 4372 old_subword, old_long, new_int, new_long); 4373 4374 // Clean up any remaining occurrences of zmem: 4375 remove_extra_zeroes(); 4376 } 4377 4378 // Explore forward from in(start) to find the first fully initialized 4379 // word, and return its offset. Skip groups of subword stores which 4380 // together initialize full words. If in(start) is itself part of a 4381 // fully initialized word, return the offset of in(start). If there 4382 // are no following full-word stores, or if something is fishy, return 4383 // a negative value. 4384 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 4385 int int_map = 0; 4386 intptr_t int_map_off = 0; 4387 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 4388 4389 for (uint i = start, limit = req(); i < limit; i++) { 4390 Node* st = in(i); 4391 4392 intptr_t st_off = get_store_offset(st, phase); 4393 if (st_off < 0) break; // return conservative answer 4394 4395 int st_size = st->as_Store()->memory_size(); 4396 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 4397 return st_off; // we found a complete word init 4398 } 4399 4400 // update the map: 4401 4402 intptr_t this_int_off = align_down(st_off, BytesPerInt); 4403 if (this_int_off != int_map_off) { 4404 // reset the map: 4405 int_map = 0; 4406 int_map_off = this_int_off; 4407 } 4408 4409 int subword_off = st_off - this_int_off; 4410 int_map |= right_n_bits(st_size) << subword_off; 4411 if ((int_map & FULL_MAP) == FULL_MAP) { 4412 return this_int_off; // we found a complete word init 4413 } 4414 4415 // Did this store hit or cross the word boundary? 4416 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt); 4417 if (next_int_off == this_int_off + BytesPerInt) { 4418 // We passed the current int, without fully initializing it. 4419 int_map_off = next_int_off; 4420 int_map >>= BytesPerInt; 4421 } else if (next_int_off > this_int_off + BytesPerInt) { 4422 // We passed the current and next int. 4423 return this_int_off + BytesPerInt; 4424 } 4425 } 4426 4427 return -1; 4428 } 4429 4430 4431 // Called when the associated AllocateNode is expanded into CFG. 4432 // At this point, we may perform additional optimizations. 4433 // Linearize the stores by ascending offset, to make memory 4434 // activity as coherent as possible. 4435 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 4436 intptr_t header_size, 4437 Node* size_in_bytes, 4438 PhaseIterGVN* phase) { 4439 assert(!is_complete(), "not already complete"); 4440 assert(stores_are_sane(phase), ""); 4441 assert(allocation() != nullptr, "must be present"); 4442 4443 remove_extra_zeroes(); 4444 4445 if (ReduceFieldZeroing || ReduceBulkZeroing) 4446 // reduce instruction count for common initialization patterns 4447 coalesce_subword_stores(header_size, size_in_bytes, phase); 4448 4449 Node* zmem = zero_memory(); // initially zero memory state 4450 Node* inits = zmem; // accumulating a linearized chain of inits 4451 #ifdef ASSERT 4452 intptr_t first_offset = allocation()->minimum_header_size(); 4453 intptr_t last_init_off = first_offset; // previous init offset 4454 intptr_t last_init_end = first_offset; // previous init offset+size 4455 intptr_t last_tile_end = first_offset; // previous tile offset+size 4456 #endif 4457 intptr_t zeroes_done = header_size; 4458 4459 bool do_zeroing = true; // we might give up if inits are very sparse 4460 int big_init_gaps = 0; // how many large gaps have we seen? 4461 4462 if (UseTLAB && ZeroTLAB) do_zeroing = false; 4463 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 4464 4465 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 4466 Node* st = in(i); 4467 intptr_t st_off = get_store_offset(st, phase); 4468 if (st_off < 0) 4469 break; // unknown junk in the inits 4470 if (st->in(MemNode::Memory) != zmem) 4471 break; // complicated store chains somehow in list 4472 4473 int st_size = st->as_Store()->memory_size(); 4474 intptr_t next_init_off = st_off + st_size; 4475 4476 if (do_zeroing && zeroes_done < next_init_off) { 4477 // See if this store needs a zero before it or under it. 4478 intptr_t zeroes_needed = st_off; 4479 4480 if (st_size < BytesPerInt) { 4481 // Look for subword stores which only partially initialize words. 4482 // If we find some, we must lay down some word-level zeroes first, 4483 // underneath the subword stores. 4484 // 4485 // Examples: 4486 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 4487 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 4488 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 4489 // 4490 // Note: coalesce_subword_stores may have already done this, 4491 // if it was prompted by constant non-zero subword initializers. 4492 // But this case can still arise with non-constant stores. 4493 4494 intptr_t next_full_store = find_next_fullword_store(i, phase); 4495 4496 // In the examples above: 4497 // in(i) p q r s x y z 4498 // st_off 12 13 14 15 12 13 14 4499 // st_size 1 1 1 1 1 1 1 4500 // next_full_s. 12 16 16 16 16 16 16 4501 // z's_done 12 16 16 16 12 16 12 4502 // z's_needed 12 16 16 16 16 16 16 4503 // zsize 0 0 0 0 4 0 4 4504 if (next_full_store < 0) { 4505 // Conservative tack: Zero to end of current word. 4506 zeroes_needed = align_up(zeroes_needed, BytesPerInt); 4507 } else { 4508 // Zero to beginning of next fully initialized word. 4509 // Or, don't zero at all, if we are already in that word. 4510 assert(next_full_store >= zeroes_needed, "must go forward"); 4511 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 4512 zeroes_needed = next_full_store; 4513 } 4514 } 4515 4516 if (zeroes_needed > zeroes_done) { 4517 intptr_t zsize = zeroes_needed - zeroes_done; 4518 // Do some incremental zeroing on rawmem, in parallel with inits. 4519 zeroes_done = align_down(zeroes_done, BytesPerInt); 4520 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4521 zeroes_done, zeroes_needed, 4522 phase); 4523 zeroes_done = zeroes_needed; 4524 if (zsize > InitArrayShortSize && ++big_init_gaps > 2) 4525 do_zeroing = false; // leave the hole, next time 4526 } 4527 } 4528 4529 // Collect the store and move on: 4530 phase->replace_input_of(st, MemNode::Memory, inits); 4531 inits = st; // put it on the linearized chain 4532 set_req(i, zmem); // unhook from previous position 4533 4534 if (zeroes_done == st_off) 4535 zeroes_done = next_init_off; 4536 4537 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 4538 4539 #ifdef ASSERT 4540 // Various order invariants. Weaker than stores_are_sane because 4541 // a large constant tile can be filled in by smaller non-constant stores. 4542 assert(st_off >= last_init_off, "inits do not reverse"); 4543 last_init_off = st_off; 4544 const Type* val = nullptr; 4545 if (st_size >= BytesPerInt && 4546 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 4547 (int)val->basic_type() < (int)T_OBJECT) { 4548 assert(st_off >= last_tile_end, "tiles do not overlap"); 4549 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 4550 last_tile_end = MAX2(last_tile_end, next_init_off); 4551 } else { 4552 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong); 4553 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 4554 assert(st_off >= last_init_end, "inits do not overlap"); 4555 last_init_end = next_init_off; // it's a non-tile 4556 } 4557 #endif //ASSERT 4558 } 4559 4560 remove_extra_zeroes(); // clear out all the zmems left over 4561 add_req(inits); 4562 4563 if (!(UseTLAB && ZeroTLAB)) { 4564 // If anything remains to be zeroed, zero it all now. 4565 zeroes_done = align_down(zeroes_done, BytesPerInt); 4566 // if it is the last unused 4 bytes of an instance, forget about it 4567 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 4568 if (zeroes_done + BytesPerLong >= size_limit) { 4569 AllocateNode* alloc = allocation(); 4570 assert(alloc != nullptr, "must be present"); 4571 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) { 4572 Node* klass_node = alloc->in(AllocateNode::KlassNode); 4573 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass(); 4574 if (zeroes_done == k->layout_helper()) 4575 zeroes_done = size_limit; 4576 } 4577 } 4578 if (zeroes_done < size_limit) { 4579 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 4580 zeroes_done, size_in_bytes, phase); 4581 } 4582 } 4583 4584 set_complete(phase); 4585 return rawmem; 4586 } 4587 4588 4589 #ifdef ASSERT 4590 bool InitializeNode::stores_are_sane(PhaseValues* phase) { 4591 if (is_complete()) 4592 return true; // stores could be anything at this point 4593 assert(allocation() != nullptr, "must be present"); 4594 intptr_t last_off = allocation()->minimum_header_size(); 4595 for (uint i = InitializeNode::RawStores; i < req(); i++) { 4596 Node* st = in(i); 4597 intptr_t st_off = get_store_offset(st, phase); 4598 if (st_off < 0) continue; // ignore dead garbage 4599 if (last_off > st_off) { 4600 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 4601 this->dump(2); 4602 assert(false, "ascending store offsets"); 4603 return false; 4604 } 4605 last_off = st_off + st->as_Store()->memory_size(); 4606 } 4607 return true; 4608 } 4609 #endif //ASSERT 4610 4611 4612 4613 4614 //============================MergeMemNode===================================== 4615 // 4616 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4617 // contributing store or call operations. Each contributor provides the memory 4618 // state for a particular "alias type" (see Compile::alias_type). For example, 4619 // if a MergeMem has an input X for alias category #6, then any memory reference 4620 // to alias category #6 may use X as its memory state input, as an exact equivalent 4621 // to using the MergeMem as a whole. 4622 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4623 // 4624 // (Here, the <N> notation gives the index of the relevant adr_type.) 4625 // 4626 // In one special case (and more cases in the future), alias categories overlap. 4627 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4628 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4629 // it is exactly equivalent to that state W: 4630 // MergeMem(<Bot>: W) <==> W 4631 // 4632 // Usually, the merge has more than one input. In that case, where inputs 4633 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4634 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4635 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4636 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4637 // 4638 // A merge can take a "wide" memory state as one of its narrow inputs. 4639 // This simply means that the merge observes out only the relevant parts of 4640 // the wide input. That is, wide memory states arriving at narrow merge inputs 4641 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4642 // 4643 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4644 // and that memory slices "leak through": 4645 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4646 // 4647 // But, in such a cascade, repeated memory slices can "block the leak": 4648 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4649 // 4650 // In the last example, Y is not part of the combined memory state of the 4651 // outermost MergeMem. The system must, of course, prevent unschedulable 4652 // memory states from arising, so you can be sure that the state Y is somehow 4653 // a precursor to state Y'. 4654 // 4655 // 4656 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4657 // of each MergeMemNode array are exactly the numerical alias indexes, including 4658 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4659 // Compile::alias_type (and kin) produce and manage these indexes. 4660 // 4661 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4662 // (Note that this provides quick access to the top node inside MergeMem methods, 4663 // without the need to reach out via TLS to Compile::current.) 4664 // 4665 // As a consequence of what was just described, a MergeMem that represents a full 4666 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4667 // containing all alias categories. 4668 // 4669 // MergeMem nodes never (?) have control inputs, so in(0) is null. 4670 // 4671 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4672 // a memory state for the alias type <N>, or else the top node, meaning that 4673 // there is no particular input for that alias type. Note that the length of 4674 // a MergeMem is variable, and may be extended at any time to accommodate new 4675 // memory states at larger alias indexes. When merges grow, they are of course 4676 // filled with "top" in the unused in() positions. 4677 // 4678 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4679 // (Top was chosen because it works smoothly with passes like GCM.) 4680 // 4681 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4682 // the type of random VM bits like TLS references.) Since it is always the 4683 // first non-Bot memory slice, some low-level loops use it to initialize an 4684 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4685 // 4686 // 4687 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4688 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4689 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4690 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4691 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4692 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4693 // 4694 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4695 // really that different from the other memory inputs. An abbreviation called 4696 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4697 // 4698 // 4699 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4700 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4701 // that "emerges though" the base memory will be marked as excluding the alias types 4702 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4703 // 4704 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4705 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4706 // 4707 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4708 // (It is currently unimplemented.) As you can see, the resulting merge is 4709 // actually a disjoint union of memory states, rather than an overlay. 4710 // 4711 4712 //------------------------------MergeMemNode----------------------------------- 4713 Node* MergeMemNode::make_empty_memory() { 4714 Node* empty_memory = (Node*) Compile::current()->top(); 4715 assert(empty_memory->is_top(), "correct sentinel identity"); 4716 return empty_memory; 4717 } 4718 4719 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4720 init_class_id(Class_MergeMem); 4721 // all inputs are nullified in Node::Node(int) 4722 // set_input(0, nullptr); // no control input 4723 4724 // Initialize the edges uniformly to top, for starters. 4725 Node* empty_mem = make_empty_memory(); 4726 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4727 init_req(i,empty_mem); 4728 } 4729 assert(empty_memory() == empty_mem, ""); 4730 4731 if( new_base != nullptr && new_base->is_MergeMem() ) { 4732 MergeMemNode* mdef = new_base->as_MergeMem(); 4733 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4734 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4735 mms.set_memory(mms.memory2()); 4736 } 4737 assert(base_memory() == mdef->base_memory(), ""); 4738 } else { 4739 set_base_memory(new_base); 4740 } 4741 } 4742 4743 // Make a new, untransformed MergeMem with the same base as 'mem'. 4744 // If mem is itself a MergeMem, populate the result with the same edges. 4745 MergeMemNode* MergeMemNode::make(Node* mem) { 4746 return new MergeMemNode(mem); 4747 } 4748 4749 //------------------------------cmp-------------------------------------------- 4750 uint MergeMemNode::hash() const { return NO_HASH; } 4751 bool MergeMemNode::cmp( const Node &n ) const { 4752 return (&n == this); // Always fail except on self 4753 } 4754 4755 //------------------------------Identity--------------------------------------- 4756 Node* MergeMemNode::Identity(PhaseGVN* phase) { 4757 // Identity if this merge point does not record any interesting memory 4758 // disambiguations. 4759 Node* base_mem = base_memory(); 4760 Node* empty_mem = empty_memory(); 4761 if (base_mem != empty_mem) { // Memory path is not dead? 4762 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4763 Node* mem = in(i); 4764 if (mem != empty_mem && mem != base_mem) { 4765 return this; // Many memory splits; no change 4766 } 4767 } 4768 } 4769 return base_mem; // No memory splits; ID on the one true input 4770 } 4771 4772 //------------------------------Ideal------------------------------------------ 4773 // This method is invoked recursively on chains of MergeMem nodes 4774 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4775 // Remove chain'd MergeMems 4776 // 4777 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4778 // relative to the "in(Bot)". Since we are patching both at the same time, 4779 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4780 // but rewrite each "in(i)" relative to the new "in(Bot)". 4781 Node *progress = nullptr; 4782 4783 4784 Node* old_base = base_memory(); 4785 Node* empty_mem = empty_memory(); 4786 if (old_base == empty_mem) 4787 return nullptr; // Dead memory path. 4788 4789 MergeMemNode* old_mbase; 4790 if (old_base != nullptr && old_base->is_MergeMem()) 4791 old_mbase = old_base->as_MergeMem(); 4792 else 4793 old_mbase = nullptr; 4794 Node* new_base = old_base; 4795 4796 // simplify stacked MergeMems in base memory 4797 if (old_mbase) new_base = old_mbase->base_memory(); 4798 4799 // the base memory might contribute new slices beyond my req() 4800 if (old_mbase) grow_to_match(old_mbase); 4801 4802 // Note: We do not call verify_sparse on entry, because inputs 4803 // can normalize to the base_memory via subsume_node or similar 4804 // mechanisms. This method repairs that damage. 4805 4806 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4807 4808 // Look at each slice. 4809 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4810 Node* old_in = in(i); 4811 // calculate the old memory value 4812 Node* old_mem = old_in; 4813 if (old_mem == empty_mem) old_mem = old_base; 4814 assert(old_mem == memory_at(i), ""); 4815 4816 // maybe update (reslice) the old memory value 4817 4818 // simplify stacked MergeMems 4819 Node* new_mem = old_mem; 4820 MergeMemNode* old_mmem; 4821 if (old_mem != nullptr && old_mem->is_MergeMem()) 4822 old_mmem = old_mem->as_MergeMem(); 4823 else 4824 old_mmem = nullptr; 4825 if (old_mmem == this) { 4826 // This can happen if loops break up and safepoints disappear. 4827 // A merge of BotPtr (default) with a RawPtr memory derived from a 4828 // safepoint can be rewritten to a merge of the same BotPtr with 4829 // the BotPtr phi coming into the loop. If that phi disappears 4830 // also, we can end up with a self-loop of the mergemem. 4831 // In general, if loops degenerate and memory effects disappear, 4832 // a mergemem can be left looking at itself. This simply means 4833 // that the mergemem's default should be used, since there is 4834 // no longer any apparent effect on this slice. 4835 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4836 // from start. Update the input to TOP. 4837 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4838 } 4839 else if (old_mmem != nullptr) { 4840 new_mem = old_mmem->memory_at(i); 4841 } 4842 // else preceding memory was not a MergeMem 4843 4844 // maybe store down a new value 4845 Node* new_in = new_mem; 4846 if (new_in == new_base) new_in = empty_mem; 4847 4848 if (new_in != old_in) { 4849 // Warning: Do not combine this "if" with the previous "if" 4850 // A memory slice might have be be rewritten even if it is semantically 4851 // unchanged, if the base_memory value has changed. 4852 set_req_X(i, new_in, phase); 4853 progress = this; // Report progress 4854 } 4855 } 4856 4857 if (new_base != old_base) { 4858 set_req_X(Compile::AliasIdxBot, new_base, phase); 4859 // Don't use set_base_memory(new_base), because we need to update du. 4860 assert(base_memory() == new_base, ""); 4861 progress = this; 4862 } 4863 4864 if( base_memory() == this ) { 4865 // a self cycle indicates this memory path is dead 4866 set_req(Compile::AliasIdxBot, empty_mem); 4867 } 4868 4869 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4870 // Recursion must occur after the self cycle check above 4871 if( base_memory()->is_MergeMem() ) { 4872 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4873 Node *m = phase->transform(new_mbase); // Rollup any cycles 4874 if( m != nullptr && 4875 (m->is_top() || 4876 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) { 4877 // propagate rollup of dead cycle to self 4878 set_req(Compile::AliasIdxBot, empty_mem); 4879 } 4880 } 4881 4882 if( base_memory() == empty_mem ) { 4883 progress = this; 4884 // Cut inputs during Parse phase only. 4885 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4886 if( !can_reshape ) { 4887 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4888 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4889 } 4890 } 4891 } 4892 4893 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4894 // Check if PhiNode::Ideal's "Split phis through memory merges" 4895 // transform should be attempted. Look for this->phi->this cycle. 4896 uint merge_width = req(); 4897 if (merge_width > Compile::AliasIdxRaw) { 4898 PhiNode* phi = base_memory()->as_Phi(); 4899 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4900 if (phi->in(i) == this) { 4901 phase->is_IterGVN()->_worklist.push(phi); 4902 break; 4903 } 4904 } 4905 } 4906 } 4907 4908 assert(progress || verify_sparse(), "please, no dups of base"); 4909 return progress; 4910 } 4911 4912 //-------------------------set_base_memory------------------------------------- 4913 void MergeMemNode::set_base_memory(Node *new_base) { 4914 Node* empty_mem = empty_memory(); 4915 set_req(Compile::AliasIdxBot, new_base); 4916 assert(memory_at(req()) == new_base, "must set default memory"); 4917 // Clear out other occurrences of new_base: 4918 if (new_base != empty_mem) { 4919 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4920 if (in(i) == new_base) set_req(i, empty_mem); 4921 } 4922 } 4923 } 4924 4925 //------------------------------out_RegMask------------------------------------ 4926 const RegMask &MergeMemNode::out_RegMask() const { 4927 return RegMask::Empty; 4928 } 4929 4930 //------------------------------dump_spec-------------------------------------- 4931 #ifndef PRODUCT 4932 void MergeMemNode::dump_spec(outputStream *st) const { 4933 st->print(" {"); 4934 Node* base_mem = base_memory(); 4935 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4936 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem; 4937 if (mem == base_mem) { st->print(" -"); continue; } 4938 st->print( " N%d:", mem->_idx ); 4939 Compile::current()->get_adr_type(i)->dump_on(st); 4940 } 4941 st->print(" }"); 4942 } 4943 #endif // !PRODUCT 4944 4945 4946 #ifdef ASSERT 4947 static bool might_be_same(Node* a, Node* b) { 4948 if (a == b) return true; 4949 if (!(a->is_Phi() || b->is_Phi())) return false; 4950 // phis shift around during optimization 4951 return true; // pretty stupid... 4952 } 4953 4954 // verify a narrow slice (either incoming or outgoing) 4955 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4956 if (!VerifyAliases) return; // don't bother to verify unless requested 4957 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error 4958 if (Node::in_dump()) return; // muzzle asserts when printing 4959 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4960 assert(n != nullptr, ""); 4961 // Elide intervening MergeMem's 4962 while (n->is_MergeMem()) { 4963 n = n->as_MergeMem()->memory_at(alias_idx); 4964 } 4965 Compile* C = Compile::current(); 4966 const TypePtr* n_adr_type = n->adr_type(); 4967 if (n == m->empty_memory()) { 4968 // Implicit copy of base_memory() 4969 } else if (n_adr_type != TypePtr::BOTTOM) { 4970 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type"); 4971 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4972 } else { 4973 // A few places like make_runtime_call "know" that VM calls are narrow, 4974 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4975 bool expected_wide_mem = false; 4976 if (n == m->base_memory()) { 4977 expected_wide_mem = true; 4978 } else if (alias_idx == Compile::AliasIdxRaw || 4979 n == m->memory_at(Compile::AliasIdxRaw)) { 4980 expected_wide_mem = true; 4981 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4982 // memory can "leak through" calls on channels that 4983 // are write-once. Allow this also. 4984 expected_wide_mem = true; 4985 } 4986 assert(expected_wide_mem, "expected narrow slice replacement"); 4987 } 4988 } 4989 #else // !ASSERT 4990 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4991 #endif 4992 4993 4994 //-----------------------------memory_at--------------------------------------- 4995 Node* MergeMemNode::memory_at(uint alias_idx) const { 4996 assert(alias_idx >= Compile::AliasIdxRaw || 4997 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()), 4998 "must avoid base_memory and AliasIdxTop"); 4999 5000 // Otherwise, it is a narrow slice. 5001 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 5002 if (is_empty_memory(n)) { 5003 // the array is sparse; empty slots are the "top" node 5004 n = base_memory(); 5005 assert(Node::in_dump() 5006 || n == nullptr || n->bottom_type() == Type::TOP 5007 || n->adr_type() == nullptr // address is TOP 5008 || n->adr_type() == TypePtr::BOTTOM 5009 || n->adr_type() == TypeRawPtr::BOTTOM 5010 || !Compile::current()->do_aliasing(), 5011 "must be a wide memory"); 5012 // do_aliasing == false if we are organizing the memory states manually. 5013 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 5014 } else { 5015 // make sure the stored slice is sane 5016 #ifdef ASSERT 5017 if (VMError::is_error_reported() || Node::in_dump()) { 5018 } else if (might_be_same(n, base_memory())) { 5019 // Give it a pass: It is a mostly harmless repetition of the base. 5020 // This can arise normally from node subsumption during optimization. 5021 } else { 5022 verify_memory_slice(this, alias_idx, n); 5023 } 5024 #endif 5025 } 5026 return n; 5027 } 5028 5029 //---------------------------set_memory_at------------------------------------- 5030 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 5031 verify_memory_slice(this, alias_idx, n); 5032 Node* empty_mem = empty_memory(); 5033 if (n == base_memory()) n = empty_mem; // collapse default 5034 uint need_req = alias_idx+1; 5035 if (req() < need_req) { 5036 if (n == empty_mem) return; // already the default, so do not grow me 5037 // grow the sparse array 5038 do { 5039 add_req(empty_mem); 5040 } while (req() < need_req); 5041 } 5042 set_req( alias_idx, n ); 5043 } 5044 5045 5046 5047 //--------------------------iteration_setup------------------------------------ 5048 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 5049 if (other != nullptr) { 5050 grow_to_match(other); 5051 // invariant: the finite support of mm2 is within mm->req() 5052 #ifdef ASSERT 5053 for (uint i = req(); i < other->req(); i++) { 5054 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 5055 } 5056 #endif 5057 } 5058 // Replace spurious copies of base_memory by top. 5059 Node* base_mem = base_memory(); 5060 if (base_mem != nullptr && !base_mem->is_top()) { 5061 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 5062 if (in(i) == base_mem) 5063 set_req(i, empty_memory()); 5064 } 5065 } 5066 } 5067 5068 //---------------------------grow_to_match------------------------------------- 5069 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 5070 Node* empty_mem = empty_memory(); 5071 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 5072 // look for the finite support of the other memory 5073 for (uint i = other->req(); --i >= req(); ) { 5074 if (other->in(i) != empty_mem) { 5075 uint new_len = i+1; 5076 while (req() < new_len) add_req(empty_mem); 5077 break; 5078 } 5079 } 5080 } 5081 5082 //---------------------------verify_sparse------------------------------------- 5083 #ifndef PRODUCT 5084 bool MergeMemNode::verify_sparse() const { 5085 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 5086 Node* base_mem = base_memory(); 5087 // The following can happen in degenerate cases, since empty==top. 5088 if (is_empty_memory(base_mem)) return true; 5089 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 5090 assert(in(i) != nullptr, "sane slice"); 5091 if (in(i) == base_mem) return false; // should have been the sentinel value! 5092 } 5093 return true; 5094 } 5095 5096 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 5097 Node* n; 5098 n = mm->in(idx); 5099 if (mem == n) return true; // might be empty_memory() 5100 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 5101 if (mem == n) return true; 5102 return false; 5103 } 5104 #endif // !PRODUCT