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